Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+resource/photo.gif filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,3 @@
+result
+**__pycache__
+.gradio

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 FoundationVision
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,12 +1,90 @@
 ---
 title: 1Prompt1Story
-emoji: 👁
-colorFrom: indigo
-colorTo: indigo
-sdk: gradio
-sdk_version: 5.13.0
 app_file: app.py
-pinned: false
+sdk: gradio
+sdk_version: 4.44.1
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+<h1 align="center">
+  <!-- <br>
+  <a href="http://www.amitmerchant.com/electron-markdownify"><img src="https://raw.githubusercontent.com/amitmerchant1990/electron-markdownify/master/app/img/markdownify.png" alt="Markdownify" width="200"></a>
+  <br> -->
+  🔥 One-Prompt-One-Story: Free-Lunch Consistent Text-to-Image Generation Using a Single Prompt
+  <br>
+</h1>
+
+
+<div align="center">
+
+[![demo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Online_Demo-blue)]()&nbsp;
+[![arXiv](https://img.shields.io/badge/arXiv%20paper-2406.06525-b31b1b.svg)]()&nbsp;
+[![project page](https://img.shields.io/badge/Project_page-More_visualizations-green)]()&nbsp;
+
+</div>
+
+
+
+<p align="center">
+  <a href="#key-features">Key Features</a> •
+  <a href="#how-to-use">How To Use</a> •
+  <a href="#license">License</a> •
+  <a href="#Citation">Citation</a> 
+</p>
+
+<p align="center">
+  <img src="./resource/photo.gif" alt="screenshot" />
+</p>
+
+
+## Key Features
+
+* Consistent Identity Image Generation.
+* Gradio Demo.
+* Consistory+ Benchmark: contains 200 prompt sets, with each set containing between 5 and 10 prompts, categorized into 8 superclasses: humans, animals, fantasy, inanimate, fairy tales, nature, technology.
+* Benchmark Generation Code.
+
+
+## How To Use
+
+To clone and run this application, you'll need [Git](https://git-scm.com) and [Node.js](https://nodejs.org/en/download/) (which comes with [npm](http://npmjs.com)) installed on your computer. From your command line:
+
+```bash
+# Clone this repository
+$ git clone https://github.com/byliutao/1Prompt1Story
+
+# Go into the repository
+$ cd 1Prompt1Story
+
+### Install dependencies ###
+$ conda create --name 1p1s python=3.10
+$ conda activate 1p1s
+# choose the right cuda version of your device
+$ conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia 
+$ conda install conda-forge::transformers 
+$ conda install -c conda-forge diffusers
+$ pip install opencv-python scipy gradio=4.44.1 sympy==1.13.1
+### Install dependencies ENDs ###
+
+# Run sample code
+$ python main.py
+
+# Run gradio demo
+$ python app.py
+
+# Run Consistory+ benchmark
+$ python -m resource.gen_benchmark --save_dir ./result/benchmark --benchmark_path ./resource/consistory+.yaml
+```
+
+> **Note**
+> If you're using Linux Bash for Windows, [see this guide](https://www.howtogeek.com/261575/how-to-run-graphical-linux-desktop-applications-from-windows-10s-bash-shell/) or use `node` from the command prompt.
+
+
+## License
+This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details.
+
+
+## Citation
+If our work assists your research, feel free to give us a star ⭐ or cite us using:
+```
+arixv
+```

app.py

@@ -0,0 +1,217 @@
+import gradio as gr
+import diffusers
+import random
+import json
+diffusers.utils.logging.set_verbosity_error()
+import torch
+from PIL import Image
+import numpy as np
+
+from unet.unet_controller import UNetController
+from main import load_unet_controller
+from unet import utils
+
+
+# Global flag to control interruption
+interrupt_flag = False
+
+
+def main_gradio(model_path, id_prompt, frame_prompt_list, precision, seed, window_length, alpha_weaken, beta_weaken, alpha_enhance, beta_enhance, ipca_drop_out, use_freeu, use_same_init_noise):
+    global interrupt_flag
+    interrupt_flag = False  # Reset the flag at the start of the function
+
+    if seed == -1:
+        seed = random.randint(0, 2**32 - 1)
+    frame_prompt_list = frame_prompt_list.split(",")
+    pipe, _ = utils.load_pipe_from_path(model_path, "cuda:1", torch.float16 if precision == "fp16" else torch.float32, precision)
+    
+    if interrupt_flag:
+        print("Generation interrupted")
+        del pipe
+        torch.cuda.empty_cache()
+
+        if 'story_image' not in locals():
+            empty_image = Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))
+            yield empty_image
+
+        return
+
+    unet_controller = load_unet_controller(pipe, "cuda:1")
+    unet_controller.Alpha_enhance = alpha_enhance
+    unet_controller.Beta_enhance = beta_enhance
+    unet_controller.Alpha_weaken = alpha_weaken
+    unet_controller.Beta_weaken = beta_weaken
+    unet_controller.Ipca_dropout = ipca_drop_out
+    unet_controller.Is_freeu_enabled = use_freeu
+    unet_controller.Use_same_init_noise = use_same_init_noise
+
+    import os
+    from datetime import datetime
+
+    current_time = datetime.now().strftime("%Y%m%d%H")
+    current_time_ = datetime.now().strftime("%M%S")
+    save_dir = os.path.join(".", f'result/{current_time}/{current_time_}_gradio_seed{seed}')
+    os.makedirs(save_dir, exist_ok=True)
+
+    generate = torch.Generator().manual_seed(seed)
+    if unet_controller.Use_ipca is True:
+        unet_controller.Store_qkv = True
+        original_prompt_embeds_mode = unet_controller.Prompt_embeds_mode
+        unet_controller.Prompt_embeds_mode = "original"
+        _ = pipe(id_prompt, generator=generate, unet_controller=unet_controller).images
+        unet_controller.Prompt_embeds_mode = original_prompt_embeds_mode
+
+    unet_controller.Store_qkv = False
+    max_window_length = utils.get_max_window_length(unet_controller, id_prompt, frame_prompt_list)
+    window_length = min(window_length, max_window_length)
+    if window_length < len(frame_prompt_list):
+        movement_lists = utils.circular_sliding_windows(frame_prompt_list, window_length)
+    else:
+        movement_lists = [movement for movement in frame_prompt_list]
+    
+    story_image_list = []
+    generate = torch.Generator().manual_seed(seed)
+    unet_controller.id_prompt = id_prompt
+    for index, movement in enumerate(frame_prompt_list):
+        if interrupt_flag:
+            print("Generation interrupted")
+            del pipe
+            torch.cuda.empty_cache()
+
+            if 'story_image' not in locals():
+                empty_image = Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))
+                yield empty_image
+
+            return
+
+        if unet_controller is not None:
+            if window_length < len(frame_prompt_list):
+                unet_controller.frame_prompt_suppress = movement_lists[index][1:]
+                unet_controller.frame_prompt_express = movement_lists[index][0]
+                gen_propmts = [f'{id_prompt} {" ".join(movement_lists[index])}']
+            else:
+                unet_controller.frame_prompt_suppress = movement_lists[:index] + movement_lists[index+1:]
+                unet_controller.frame_prompt_express = movement_lists[index]
+                gen_propmts = [f'{id_prompt} {" ".join(movement_lists)}']
+        else:
+            gen_propmts = f'{id_prompt} {movement}'
+
+
+        print(f"suppress: {unet_controller.frame_prompt_suppress}")
+        print(f"express: {unet_controller.frame_prompt_express}")
+        print(f'id_prompt: {id_prompt}')
+        print(f"gen_propmts: {gen_propmts}")
+
+        if unet_controller is not None and unet_controller.Use_same_init_noise is True:
+            generate = torch.Generator().manual_seed(seed)
+
+        images = pipe(gen_propmts, generator=generate, unet_controller=unet_controller).images
+        story_image_list.append(images[0])
+
+        story_image = np.concatenate(story_image_list, axis=1)
+        story_image = Image.fromarray(story_image.astype(np.uint8))
+
+        yield story_image
+        import os
+        images[0].save(os.path.join(save_dir, f'{id_prompt} {unet_controller.frame_prompt_express}.jpg'))
+
+    story_image.save(os.path.join(save_dir, 'story_image.jpg'))
+
+    import gc
+    del pipe
+    gc.collect()
+    torch.cuda.empty_cache()
+
+# Gradio interface
+def gradio_interface():
+    global interrupt_flag
+
+    with gr.Blocks() as demo:
+        gr.Markdown("### Consistent Image Generation with 1Prompt1Story")
+
+        # Load JSON data
+        with open('./resource/example.json', 'r') as f:
+            data = json.load(f)
+
+        # Extract id_prompts and frame_prompts
+        id_prompts = [item['id_prompt'] for item in data['combinations']]
+        frame_prompts = [", ".join(item['frame_prompt_list']) for item in data['combinations']]
+
+        # Input fields
+        id_prompt = gr.Dropdown(
+            label="ID Prompt",
+            choices=id_prompts,
+            value=id_prompts[0],
+            allow_custom_value=True
+        )
+        frame_prompt_list = gr.Dropdown(
+            label="Frame Prompts (comma-separated)",
+            choices=frame_prompts,
+            value=frame_prompts[0],
+            allow_custom_value=True
+        )
+        model_path = gr.Dropdown(
+            label="Model Path",
+            choices=["stabilityai/stable-diffusion-xl-base-1.0", "RunDiffusion/Juggernaut-X-v10", "playgroundai/playground-v2.5-1024px-aesthetic", "SG161222/RealVisXL_V4.0", "RunDiffusion/Juggernaut-XI-v11", "SG161222/RealVisXL_V5.0"],
+            value="playgroundai/playground-v2.5-1024px-aesthetic",
+            allow_custom_value=True
+        )
+        
+        with gr.Row():
+            seed = gr.Slider(label="Seed (set -1 for random seed)", minimum=-1, maximum=10000, value=-1, step=1)
+            window_length = gr.Slider(label="Window Length", minimum=1, maximum=20, value=10, step=1)
+
+        with gr.Row():
+            alpha_weaken = gr.Number(label="Alpha Weaken", value=UNetController.Alpha_weaken, interactive=True, step=0.01)
+            beta_weaken = gr.Number(label="Beta Weaken", value=UNetController.Beta_weaken, interactive=True, step=0.01)
+            alpha_enhance = gr.Number(label="Alpha Enhance", value=UNetController.Alpha_enhance, interactive=True, step=0.001)
+            beta_enhance = gr.Number(label="Beta Enhance", value=UNetController.Beta_enhance, interactive=True, step=0.1)
+        
+        with gr.Row():
+            ipca_drop_out = gr.Number(label="Ipca Dropout", value=UNetController.Ipca_dropout, interactive=True, step=0.1, minimum=0, maximum=1)
+            precision = gr.Dropdown(label="Precision", choices=["fp16", "fp32"], value="fp16")
+            use_freeu = gr.Dropdown(label="Use FreeU", choices=[False, True], value=UNetController.Is_freeu_enabled)
+            use_same_init_noise = gr.Dropdown(label="Use Same Init Noise", choices=[True, False], value=UNetController.Use_same_init_noise)
+        
+        reset_button = gr.Button("Reset to Default")
+
+        def reset_values():
+            return UNetController.Alpha_weaken, UNetController.Beta_weaken, UNetController.Alpha_enhance, UNetController.Beta_enhance, UNetController.Ipca_dropout, "fp16", UNetController.Is_freeu_enabled, UNetController.Use_same_init_noise
+
+        reset_button.click(
+            fn=reset_values,
+            inputs=[],
+            outputs=[alpha_weaken, beta_weaken, alpha_enhance, beta_enhance, ipca_drop_out, precision, use_freeu, use_same_init_noise]
+        )
+
+        # Output
+        output_gallery = gr.Image()
+
+        # Buttons
+        generate_button = gr.Button("Generate Images")
+        interrupt_button = gr.Button("Interrupt")
+
+        def interrupt_generation():
+            global interrupt_flag
+            interrupt_flag = True
+
+        interrupt_button.click(
+            fn=interrupt_generation,
+            inputs=[],
+            outputs=[]
+        )
+
+        generate_button.click(
+            fn=main_gradio,
+            inputs=[
+                model_path, id_prompt, frame_prompt_list, precision, seed, window_length, alpha_weaken, beta_weaken, alpha_enhance, beta_enhance, ipca_drop_out, use_freeu, use_same_init_noise
+            ],
+            outputs=output_gallery
+        )
+
+    return demo
+
+
+if __name__ == "__main__":
+    demo = gradio_interface()
+    demo.launch(share=True)

main.py

@@ -0,0 +1,97 @@
+import os
+import torch
+import random
+import diffusers
+import torch.utils
+import unet.utils as utils
+from unet.unet_controller import UNetController
+import argparse
+from datetime import datetime
+
+diffusers.utils.logging.set_verbosity_error()
+
+def load_unet_controller(pipe, device):
+    unet_controller = UNetController()
+    unet_controller.device = device
+    unet_controller.tokenizer = pipe.tokenizer
+
+    return unet_controller
+
+
+def generate_images(unet_controller: UNetController, pipe, id_prompt, frame_prompt_list, save_dir, window_length, seed, verbose=True):
+    generate = torch.Generator().manual_seed(seed)
+    if unet_controller.Use_ipca is True:
+        unet_controller.Store_qkv = True
+        original_prompt_embeds_mode = unet_controller.Prompt_embeds_mode
+        unet_controller.Prompt_embeds_mode = "original"
+        _ = pipe(id_prompt, generator=generate, unet_controller=unet_controller).images
+        unet_controller.Prompt_embeds_mode = original_prompt_embeds_mode
+
+
+    unet_controller.Store_qkv = False
+    images, story_image = utils.movement_gen_story_slide_windows(
+        id_prompt, frame_prompt_list, pipe, window_length, seed, unet_controller, save_dir, verbose=verbose
+    )
+
+    return images, story_image
+
+
+def main(device, model_path, save_dir, id_prompt, frame_prompt_list, precision, seed, window_length):
+    pipe, _ = utils.load_pipe_from_path(model_path, device, torch.float16 if precision == "fp16" else torch.float32, precision)
+    
+    unet_controller = load_unet_controller(pipe, device)          
+    images, story_image = generate_images(unet_controller, pipe, id_prompt, frame_prompt_list, save_dir, window_length, seed)
+
+    return images, story_image
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Generate images using a specific device.")
+    parser.add_argument('--device', type=str, default='cuda:0', help='Device to use for computation (e.g., cuda:0, cpu)')
+    parser.add_argument('--model_path', type=str, default='playgroundai/playground-v2.5-1024px-aesthetic', help='Path to the model')
+    parser.add_argument('--project_base_path', type=str, default='.', help='Path to save the generated images')
+    parser.add_argument('--id_prompt', type=str, default="A photo of a red fox with coat", help='Initial prompt for image generation')
+    parser.add_argument('--frame_prompt_list', type=str, nargs='+', default=[
+        "wearing a scarf in a meadow",
+        "playing in the snow",
+        "at the edge of a village with river",
+    ], help='List of frame prompts')
+    parser.add_argument('--precision', type=str, choices=["fp16", "fp32"], default="fp16", help='Model precision')
+    parser.add_argument('--seed', type=int, default=42, help='Random seed for generation')
+    parser.add_argument('--window_length', type=int, default=10, help='Window length for story generation')
+    parser.add_argument('--save_padding', type=str, default='test', help='Padding for save directory')
+    parser.add_argument('--random_seed', action='store_true', help='Use random seed')
+    parser.add_argument('--json_path', type=str,)
+    
+    args = parser.parse_args()
+    if args.random_seed:
+        args.seed = random.randint(0, 1000000)
+
+    current_time = datetime.now().strftime("%Y%m%d%H")
+    current_time_ = datetime.now().strftime("%M%S")
+    save_dir = os.path.join(args.project_base_path, f'result/{current_time}/{current_time_}_{args.save_padding}_seed{args.seed}')
+    os.makedirs(save_dir, exist_ok=True)
+
+    if args.json_path is None:
+        id_prompt = "A cinematic portrait of a man and a woman standing together"
+        frame_prompt_list = [
+                "under a sky full of stars",
+                "on a bustling city street at night",
+                "in a dimly lit jazz club",
+                "walking along a sandy beach at sunset",
+                "in a cozy coffee shop with large windows",
+                "in a vibrant art gallery surrounded by paintings",
+                "under an umbrella during a soft rain",
+                "on a quiet park bench amidst falling leaves",
+                "standing on a rooftop overlooking the city skyline"
+            ]
+        main(args.device, args.model_path, save_dir, id_prompt, frame_prompt_list, args.precision, args.seed, args.window_length)
+    else:
+        import json
+        with open(args.json_path, "r") as file:
+            data = json.load(file)
+
+        combinations = data["combinations"]
+
+        for combo in combinations:
+            main(args.device, args.model_path, save_dir, combo['id_prompt'], combo['frame_prompt_list'], args.precision, args.seed, args.window_length)

requirements.txt

@@ -0,0 +1 @@
+pytorch transformers diffusers

resource/consistory+.yaml

@@ -0,0 +1,1894 @@
+animals:
+- concept_token: phoenix
+  settings:
+  - rising from a fiery ashes
+  - soaring through a glowing sky
+  - perching on a mountain peak
+  - singing a haunting melody
+  - igniting flames with its wings
+  style: A fiery and majestic illustration of
+  subject: A phoenix with bright orange feathers
+- concept_token: zebra
+  settings:
+  - grazing alongside a river
+  - running in a herd across the plains
+  - resting under the shade of an acacia tree
+  - crossing a dusty path in the wild
+  - protecting its young in the savannah
+  style: A vibrant and striking portrait of
+  subject: A zebra with black and white stripes
+- concept_token: cheetah
+  settings:
+  - sprinting across the savannah
+  - stalking a gazelle in the grass
+  - relaxing in the shade under a tree
+  - marking territory with its scent
+  - napping after a successful hunt
+  style: A sleek and fast depiction of
+  subject: A cheetah with sharp eyes
+- concept_token: walrus
+  settings:
+  - lounging on an ice floe
+  - diving for clams in the sea
+  - bellowing from the shore
+  - resting near a snowy coastline
+  - swimming gracefully in the ocean
+  style: A massive and majestic depiction of
+  subject: A walrus with large tusks
+- concept_token: zebra
+  settings:
+  - grazing under the sun in the savannah
+  - running with a herd through the grasslands
+  - standing near a watering hole
+  - crossing a river in the wild
+  - trotting beside a wildebeest
+  style: A bold and striking illustration of
+  subject: A zebra with black and white stripes
+- concept_token: cheetah
+  settings:
+  - sprinting at top speed in the savannah
+  - resting in the shade of a tree
+  - hunting a gazelle in tall grass
+  - stretching before a sprint
+  - watching the horizon from a hill
+  style: A sleek and fast portrayal of
+  subject: A cheetah with spotted fur
+- concept_token: gorilla
+  settings:
+  - sitting in a jungle clearing
+  - playing with a baby gorilla
+  - searching for fruit in the trees
+  - relaxing by a waterfall
+  - beating its chest to assert dominance
+  style: A strong and wise illustration of
+  subject: A gorilla with dark fur
+- concept_token: dog
+  settings:
+  - chasing a frisbee
+  - dressed in a raincoat
+  - in a city alley
+  - jumping over a puddle
+  - at a veterinarian's office
+  - playing in a park
+  - sleeping on a couch
+  - running on a beach
+  style: A hyper-realistic digital painting of
+  subject: A dog
+- concept_token: parrot
+  settings:
+  - in a tropical rainforest
+  - singing a song
+  - on the water
+  - sitting on a rock
+  - in a rain storm
+  - perched on a branch
+  - eating fruit
+  style: A watercolor illustration of
+  subject: A Scarlet parrot with vibrant red, yellow, and blue feathers
+- concept_token: dog
+  settings:
+  - wearing a bandana
+  - on a beach
+  - in a city alley
+  - in a snowy backyard
+  - at a veterinarian's office
+  - playing with a toy
+  - barking at a squirrel
+  style: A 3D animation of
+  subject: A black and white dog with yellow collar
+- concept_token: fox
+  settings:
+  - wearing a scarf
+  - dressed in a warm coat
+  - wearing a playful bow
+  - resting in a den
+  - near a campsite
+  - exploring a forest
+  - looking at the stars
+  style: A hyper-realistic digital painting of
+  subject: A red fox with a vibrant red coat, white belly, and bushy tail
+- concept_token: puppy
+  settings:
+  - in a pet store
+  - eating his food
+  - wearing a training harness
+  - dressed in a bandana
+  - at a lake
+  - playing with a ball
+  - sleeping in a bed
+  style: A 3D animation of
+  subject: A cute Labrador puppy with a glossy, chocolate brown coat
+- concept_token: dog
+  settings:
+  - wearing a bandana
+  - biting a bone
+  - wearing a birthday hat
+  - in a snowy backyard
+  - sitting by a fireplace
+  style: A 3D animation of
+  subject: A black and white dog with yellow collar
+- concept_token: kitten
+  settings:
+  - in a garden
+  - dressed in a cute sweater
+  - wearing a collar with a bell
+  - dressed in a superhero cape
+  - running through a field
+  - playing with a toy
+  - sitting in a basket
+  style: A watercolor illustration of
+  subject: A cute kitten with sleek, cream-colored fur and striking blue eyes
+- concept_token: puppy
+  settings:
+  - in a pet store
+  - in a grassy yard
+  - wearing a training harness
+  - dressed in a bandana
+  - wearing a life vest
+  style: ''
+  subject: A puppy
+- concept_token: puppy
+  settings:
+  - wearing a small sweater
+  - digging a hole
+  - wearing a training harness
+  - sticking head out of the car window
+  - swimming
+  - playing in the yard
+  - chasing a ball
+  style: A watercolor illustration of
+  subject: A puppy
+- concept_token: cat
+  settings:
+  - playing with a yarn ball
+  - sleeping in a box
+  - wearing a fluffy collar
+  - climbing a tree
+  - sitting on a shelf
+  style: A hyper-realistic digital painting of
+  subject: A cat
+- concept_token: cat
+  settings:
+  - wearing a bow tie
+  - dressed in a Halloween costume
+  - in a busy alley
+  - on a sled
+  - wearing a small bell
+  style: A hyper-realistic digital painting of
+  subject: A cat
+- concept_token: kitten
+  settings:
+  - wearing a tiny hat
+  - on a couch
+  - in a hospital
+  - playing with a feather toy
+  - on a rooftop
+  style: ''
+  subject: A cute kitten with sleek, cream-colored fur and striking blue eyes
+- concept_token: hedgehog
+  settings:
+  - in a cozy nest
+  - dressed in a miniature jacket
+  - wearing a small collar
+  - dressed in a festive outfit
+  - wearing a flower crown
+  style: A 3D animation of
+  subject: A happy hedgehog
+- concept_token: horse
+  settings:
+  - in a stable
+  - jumping over a hurdle
+  - in a snowy field
+  - in a mountain trail
+  - in a busy street
+  style: A hyper-realistic digital painting of
+  subject: A Palomino horse with a golden coat and a flowing, white mane and tail
+fairy_tales:
+- concept_token: fairy
+  settings:
+  - fluttering among glowing fireflies
+  - sprinkling pixie dust in the air
+  - hiding inside a blooming flower
+  - weaving spells under an ancient tree
+  - guiding travelers with a magical lantern
+  style: A magical drawing of
+  subject: A delicate fairy with sparkling wings
+- concept_token: princess
+  settings:
+  - walking in a garden of enchanted roses
+  - gazing at the stars from a castle balcony
+  - singing to woodland animals in a forest
+  - wearing a shimmering gown at a royal ball
+  - seeking wisdom from an old magical mirror
+  style: A dreamy illustration of
+  subject: A beautiful princess with a kind smile
+- concept_token: troll
+  settings:
+  - lurking beneath an old stone bridge
+  - guarding a cave filled with treasures
+  - chasing intruders through dark woods
+  - sharpening a massive club near a fire
+  - watching over a mountain pass silently
+  style: A dark fantasy depiction of
+  subject: A menacing troll with rough skin
+- concept_token: witch
+  settings:
+  - stirring a bubbling potion in a cauldron
+  - flying on a broomstick under a full moon
+  - casting spells in a candlelit cabin
+  - gathering herbs in a misty forest
+  - reading incantations from an ancient book
+  style: A dark and enchanting artwork of
+  subject: A mysterious witch in flowing robes
+- concept_token: phoenix
+  settings:
+  - rising from ashes in a fiery burst
+  - soaring through the golden clouds of dawn
+  - perched atop a flaming tree
+  - lighting the dark forest with its glow
+  - circling a distant volcano in flight
+  style: A vibrant fantasy drawing of
+  subject: A majestic phoenix with flaming wings
+- concept_token: elf
+  settings:
+  - crafting a bow from enchanted wood
+  - guiding travelers through an ancient forest
+  - practicing archery under a silver moon
+  - guarding a hidden woodland village
+  - reading an ancient map by firelight
+  style: A detailed character design of
+  subject: A graceful elf with pointed ears
+- concept_token: snow queen
+  settings:
+  - commanding a storm in an icy palace
+  - creating snowflakes with a wave of her hand
+  - watching over a frozen kingdom
+  - walking through a glittering ice cave
+  - sitting on a throne of frosted crystals
+  style: A majestic winter-themed artwork of
+  subject: A regal snow queen with icy beauty
+- concept_token: goblin
+  settings:
+  - sneaking into a hidden treasure vault
+  - crafting traps in a dark cave
+  - trading stolen trinkets at a market
+  - hiding from sunlight under a tree
+  - scavenging through enchanted ruins
+  style: A mischievous fantasy depiction of
+  subject: A cunning goblin with sharp features
+- concept_token: magic book
+  settings:
+  - floating in midair, pages turning
+  - glowing faintly in a candlelit room
+  - locked with an ornate golden clasp
+  - whispering spells to its reader
+  - "lying open on a wizard\u2019s desk"
+  style: A mystical object illustration of
+  subject: An ancient magic book filled with secrets
+- concept_token: unicorn
+  settings:
+  - galloping through a dense, enchanted forest
+  - drinking from a crystalline stream
+  - standing in a meadow of blooming flowers
+  - beneath a rainbow in a serene valley
+  - resting in the shadow of ancient ruins
+  - among the clouds at dawn
+  style: A magical artwork of
+  subject: A unicorn with a gleaming silver horn
+- concept_token: fairy
+  settings:
+  - hovering over a moonlit pond
+  - dancing on the petals of a giant flower
+  - hiding in the hollow of an ancient tree
+  - spreading fairy dust over a sleeping village
+  - sitting on a mushroom in a magical forest
+  - playing with fireflies at dusk
+  - weaving through a field of wildflowers
+  style: A whimsical painting of
+  subject: A delicate fairy with translucent wings
+- concept_token: wizard
+  settings:
+  - in a tower filled with ancient tomes and artifacts
+  - casting a spell by the light of a full moon
+  - standing before a magical portal in the forest
+  - summoning a storm over a mountain peak
+  - writing runes in a dusty spellbook
+  - mixing potions in a dimly lit chamber
+  - consulting a crystal ball
+  style: A mystical illustration of
+  subject: A wise wizard with a long, flowing beard
+- concept_token: griffin
+  settings:
+  - soaring over a sprawling desert
+  - perched on a high cliff watching the horizon
+  - standing guard over a hidden treasure
+  - flying above a lush green valley
+  - resting on a rocky outcrop at sunset
+  - gliding through the clouds in a clear sky
+  - hunting in a dense forest
+  style: A majestic painting of
+  subject: A majestic griffin with golden feathers
+- concept_token: centaur
+  settings:
+  - galloping through a grassy plain at dawn
+  - patrolling the edge of a dense forest
+  - leading a group of warriors through a mountain pass
+  - beside a river under a twilight sky
+  - practicing archery in a secluded glade
+  - at the foot of a towering cliff
+  - under a sky filled with storm clouds
+  style: An adventurous illustration of
+  subject: A powerful centaur with a bow
+- concept_token: nymph
+  settings:
+  - dancing on a moonlit pond
+  - singing beside a sparkling waterfall
+  - playing with fish in a crystal-clear stream
+  - resting among blooming lilies
+  - under a weeping willow at dawn
+  - in a serene glade surrounded by flowers
+  - floating in a tranquil lake
+  - gathering dew at the break of day
+  style: A serene painting of
+  subject: A delicate water nymph with flowing hair
+- concept_token: troll
+  settings:
+  - under a stone bridge covered in ivy
+  - guarding a treasure chest in a dark cave
+  - helping travelers across a river
+  - sitting by a campfire in a foggy forest
+  - building a shelter from fallen logs
+  - fishing in a quiet stream at dusk
+  - carving runes into a rock
+  - resting under a large oak tree
+  style: A heartwarming illustration of
+  subject: A friendly troll with moss-covered skin
+- concept_token: ogre
+  settings:
+  - living in a cozy cave in the forest
+  - fishing in a quiet lake under a cloudy sky
+  - helping lost travelers find their way
+  - gathering berries in a sunlit meadow
+  - sitting by a roaring campfire
+  - playing with forest animals in a glade
+  - building a shelter from fallen logs
+  - carrying a bundle of firewood through the forest
+  style: A friendly depiction of
+  subject: A gentle ogre with a broad smile
+- concept_token: dwarf
+  settings:
+  - mining for gems in a glittering cave
+  - crafting weapons at a forge
+  - drinking ale in a bustling tavern
+  - exploring a network of underground tunnels
+  - climbing a steep mountain path
+  - gathering herbs in a forest clearing
+  - carving runes into stone tablets
+  - resting by a roaring fireplace
+  style: A robust painting of
+  subject: A sturdy dwarf with a thick beard
+- concept_token: goblin
+  settings:
+  - sneaking through a dark market at midnight
+  - playing tricks on travelers in a village
+  - hiding in the shadows of a narrow alleyway
+  - stealing shiny objects from a merchant's stall
+  - scurrying through a dense forest at twilight
+  - climbing the walls of a deserted castle
+  - laughing by a flickering campfire
+  - running through an underground tunnel
+  style: A quirky artwork of
+  subject: A mischievous goblin with sharp features
+- concept_token: witch
+  settings:
+  - brewing a potion in a bubbling cauldron
+  - casting a spell in a dark, enchanted forest
+  - flying on a broomstick under a full moon
+  - preparing charms in a cluttered cottage
+  - consulting an ancient grimoire by candlelight
+  - in a shadowy glade surrounded by glowing eyes
+  - watching over a bubbling potion in a misty cave
+  style: A mysterious painting of
+  subject: A cunning witch with a pointy hat
+- concept_token: satyr
+  settings:
+  - dancing in a moonlit glade
+  - playing tunes beside a babbling brook
+  - hiding behind a tree in a sun-dappled forest
+  - leading a merry chase through a meadow
+  - resting on a boulder in a twilight grove
+  - frolicking among wildflowers
+  - drinking from a natural spring
+  - singing under the light of the full moon
+  style: A playful painting of
+  subject: A mischievous satyr with a reed pipe
+- concept_token: dryad
+  settings:
+  - resting against an ancient oak tree
+  - walking through a sunlit grove
+  - dancing with the wind in a forest clearing
+  - watching over a grove of saplings
+  - playing with woodland creatures in the morning mist
+  - singing to the birds in the early dawn light
+  - weaving flowers into her hair
+  - resting in the shade of a grand tree
+  style: A naturalistic painting of
+  subject: A serene dryad with leafy hair
+- concept_token: princess
+  settings:
+  - walking in a garden of enchanted roses
+  - gazing at the stars from a castle balcony
+  - singing to woodland animals in a forest
+  - wearing a shimmering gown at a royal ball
+  - seeking wisdom from an old magical mirror
+  style: A dreamy illustration of
+  subject: A beautiful princess
+- concept_token: wizard
+  settings:
+  - casting spells in a mystical tower
+  - studying ancient tomes in a library
+  - summoning creatures from a cauldron
+  - battling a dragon in a dark forest
+  - creating potions under a full moon
+  style: A mystical painting of
+  subject: A wise wizard
+- concept_token: witch
+  settings:
+  - stirring a potion in a bubbling cauldron
+  - flying on a broom under a full moon
+  - reading from a spellbook in a dark hut
+  - casting a curse in a misty forest
+  - collecting herbs in a hidden glade
+  style: A dark and eerie drawing of
+  subject: A mysterious witch
+- concept_token: troll
+  settings:
+  - guarding a stone bridge in the mountains
+  - sitting by a campfire in the woods
+  - threatening travelers in a dark cave
+  - wandering through a foggy swamp
+  - grumbling in an underground lair
+  style: A rugged drawing of
+  subject: A grumpy troll
+- concept_token: fairy godmother
+  settings:
+  - waving a wand to transform pumpkins
+  - granting wishes in a sparkling forest
+  - offering advice to a princess
+  - preparing magical gifts in a cozy home
+  - visiting a poor family with a blessing
+  style: A heartwarming illustration of
+  subject: A kind fairy godmother
+- concept_token: ogre
+  settings:
+  - stomping through a village in a rage
+  - hiding in a dark cave
+  - eating a giant feast by a fire
+  - chasing heroes through the forest
+  - resting by a muddy river
+  style: A frightening drawing of
+  subject: A fearsome ogre
+- concept_token: gnome
+  settings:
+  - working in a hidden underground garden
+  - sitting by a cozy fireplace with a pipe
+  - traveling with a mushroom cart
+  - tending to the plants in a forest
+  - playing tricks in a quiet village
+  style: A charming sketch of
+  subject: A friendly gnome
+- concept_token: goblin
+  settings:
+  - sneaking through a dark alley
+  - working in a dank forge
+  - hoarding treasure in a hidden cave
+  - stealing from travelers in the night
+  - brewing potions in a dark hut
+  style: A sneaky illustration of
+  subject: A cunning goblin
+- concept_token: faun
+  settings:
+  - playing a flute by a bubbling stream
+  - dancing in a circle of mushrooms
+  - guiding travelers through a magical forest
+  - leading a procession of woodland creatures
+  - sitting by a fire telling stories
+  style: A peaceful sketch of
+  subject: A playful faun
+- concept_token: phoenix
+  settings:
+  - rising from the ashes in a burst of fire
+  - flying through a fiery sky
+  - resting on a burning tree
+  - appearing in a flash of flame
+  - spreading its wings over a burning city
+  style: A fiery illustration of
+  subject: A majestic phoenix
+- concept_token: unicorn
+  settings:
+  - galloping through a flower field
+  - standing by a sparkling waterfall
+  - racing across a rainbow bridge
+  - protecting a sacred grove
+  - spreading magic in the forest
+  style: A colorful illustration of
+  subject: A graceful unicorn
+- concept_token: witch
+  settings:
+  - flying on a broomstick under a full moon
+  - brewing a potion in a bubbling cauldron
+  - reading spells in an ancient book
+  - preparing a magical brew in a forest
+  - cursing a lost traveler in the woods
+  style: A dark drawing of
+  subject: A mysterious witch
+- concept_token: elf
+  settings:
+  - singing songs of old in a moonlit grove
+  - crafting bows and arrows in a woodland cabin
+  - gathering herbs in an enchanted forest
+  - fighting in a battle with a gleaming sword
+  - celebrating with fellow elves by a fire
+  style: A serene painting of
+  subject: A wise elf
+- concept_token: troll
+  settings:
+  - lurking under a bridge
+  - hiding in a cave at night
+  - eating mushrooms in a dark forest
+  - grumbling as it patrols the woods
+  - guarding a treasure chest
+  style: A grumpy drawing of
+  subject: A massive troll
+- concept_token: siren
+  settings:
+  - singing to passing sailors
+  - playing an enchanted harp on a rocky shore
+  - luring ships with a beautiful melody
+  - hiding in the depths of the sea
+  - sitting on a cliff watching the sunset
+  style: A haunting painting of
+  subject: A seductive siren
+- concept_token: pegasus
+  settings:
+  - soaring over a green valley
+  - flying through the clouds above a mountain range
+  - galloping through the sky with wings spread wide
+  - racing along the edge of the ocean
+  - resting on a soft cloud at dawn
+  style: A majestic painting of
+  subject: A graceful pegasus
+- concept_token: gnome
+  settings:
+  - tinkering in an underground workshop
+  - working in a lush garden
+  - guarding the entrance to a secret cave
+  - sitting on a mushroom with a pipe
+  - creating magical trinkets for travelers
+  style: A cozy illustration of
+  subject: A friendly gnome
+- concept_token: faun
+  settings:
+  - playing a flute by a gentle stream
+  - dancing among ancient trees
+  - gathering flowers for a wreath
+  - guiding lost travelers through the forest
+  - resting under a giant oak tree
+  style: A whimsical drawing of
+  subject: A playful faun
+fantasy:
+- concept_token: griffin
+  settings:
+  - soaring above a golden desert
+  - nesting on a mountain cliff
+  - hunting in a forest at dawn
+  - standing proudly in front of a castle
+  - flying through a stormy sky
+  style: A majestic and powerful illustration of
+  subject: A griffin with the body of a lion and the wings of an eagle
+- concept_token: centaur
+  settings:
+  - running through an open meadow
+  - playing a tune on a flute
+  - practicing swordsmanship in a field
+  - sitting by a campfire at night
+  - galloping through a forest trail
+  style: A heroic nature illustration of
+  subject: A centaur with the body of a horse and the torso of a warrior
+- concept_token: mermaid
+  settings:
+  - swimming in a coral reef
+  - basking on a sunlit rock
+  - combing her hair with a shell
+  - singing in an underwater cave
+  - playing with colorful fish
+  style: A dreamy underwater illustration of
+  subject: A beautiful mermaid with a shimmering tail
+- concept_token: wizard
+  settings:
+  - casting spells in a tower
+  - reading ancient books in a library
+  - brewing potions in a dark cave
+  - summoning creatures in a circle
+  - wandering through a mystical forest
+  style: A mystical and powerful illustration of
+  subject: A wise wizard with a long, flowing beard
+- concept_token: werewolf
+  settings:
+  - howling at the full moon
+  - prowling through a misty forest
+  - transforming in the light of the moon
+  - hunting in the dark woods
+  - running on all fours under a starry sky
+  style: A terrifying and wild illustration of
+  subject: A werewolf with glowing yellow eyes
+- concept_token: ogre
+  settings:
+  - stomping through a muddy swamp
+  - sitting by a campfire roasting food
+  - swinging a massive club in battle
+  - chasing intruders in the forest
+  - resting in a cave with treasure
+  style: A brutish and strong illustration of
+  subject: A large ogre with green skin and a rough demeanor
+- concept_token: siren
+  settings:
+  - singing on a rocky cliff
+  - luring sailors to their doom
+  - swimming in the deep ocean
+  - resting on a wave-kissed shore
+  - weaving spells with her voice
+  style: A haunting and alluring illustration of
+  subject: A siren with long flowing hair and a melodious voice
+- concept_token: elf queen
+  settings:
+  - sitting on a crystal throne
+  - overseeing the elven kingdom
+  - speaking with ancient spirits
+  - leading her people into battle
+  - walking through a mystical forest
+  style: A regal and elegant illustration of
+  subject: An elf queen with a jeweled crown
+- concept_token: hobbit
+  settings:
+  - lounging by a cozy fireplace
+  - planting flowers in a garden
+  - enjoying a feast at home
+  - walking through the rolling hills
+  - smoking a pipe under a tree
+  style: A peaceful and rustic illustration of
+  subject: A hobbit with large feet and a warm smile
+- concept_token: mermaid
+  settings:
+  - swimming in a coral reef
+  - basking on a sunlit rock
+  - combing her hair with a shell
+  - singing in an underwater cave
+  - playing with colorful fish
+  style: A dreamy underwater illustration of
+  subject: A beautiful mermaid with a shimmering tail
+- concept_token: goblin
+  settings:
+  - sneaking into a treasure cave
+  - trading trinkets at a market
+  - sharpening weapons in a lair
+  - arguing over a shiny object
+  - lurking in the shadows of ruins
+  style: A dark fantasy illustration of
+  subject: A mischievous goblin with sharp teeth
+- concept_token: sorceress
+  settings:
+  - conjuring a glowing orb
+  - surrounded by enchanted flames
+  - chanting in a stone circle
+  - reading from an ancient tome
+  - standing by a bubbling cauldron
+  style: A captivating fantasy portrait of
+  subject: A powerful sorceress in elegant robes
+- concept_token: dwarf
+  settings:
+  - forging weapons in a fiery forge
+  - mining deep underground
+  - standing guard by a stone door
+  - drinking from a golden goblet
+  - carrying a heavy battleaxe
+  style: A gritty fantasy artwork of
+  subject: A stout dwarf with a long beard
+- concept_token: troll
+  settings:
+  - sitting on a mossy boulder
+  - guarding a bridge in the woods
+  - clutching a massive club
+  - chasing intruders from a cave
+  - growling in the moonlight
+  style: A dark and menacing depiction of
+  subject: A hulking troll with a fearsome appearance
+- concept_token: fairy
+  settings:
+  - wearing gossamer wings
+  - frolicking in the air
+  - wearing a crown of twigs
+  - near a shimmering brook
+  - dancing in moonlight
+  - in an enchanted forest
+  - sitting on a mushroom
+  style: A watercolor illustration of
+  subject: A fairy
+- concept_token: elf
+  settings:
+  - in an elven city of trees
+  - cleaning a sword
+  - reading an ancient text
+  - flying in the air
+  - in a library of scrolls
+  - meditating by a waterfall
+  - at a council meeting
+  style: A photo of
+  subject: An elf with long, silver hair, eyes like polished amethyst, and ears that
+    curve elegantly upwards
+- concept_token: unicorn
+  settings:
+  - prancing near the water
+  - beside a crystal lake
+  - wearing a silver bridle
+  - under a canopy of stars
+  - wearing a harness of sunbeams
+  - in a magical forest
+  style: A black and white sketch of
+  subject: A unicorn
+- concept_token: goblin
+  settings:
+  - wearing a cloak
+  - haggling over goods
+  - wearing a belt of tools
+  - dressed in a guard's uniform
+  - in a dungeon's depths
+  - scavenging for treasures
+  style: A 3D animation of
+  subject: A goblin with patchy, leathery skin and oversized ears, carrying a tiny,
+    glowing lantern
+- concept_token: fairy
+  settings:
+  - in a mystical glen
+  - atop a dew-covered flower
+  - inside a hollowed-out tree
+  - collecting morning dew
+  - under a full moon
+  - dancing in the rain
+  style: A hyper-realistic digital painting of
+  subject: A fairy
+- concept_token: leprechaun
+  settings:
+  - counting gold coins
+  - at the end of a rainbow
+  - setting a trap
+  - in an underground workshop
+  - performing a jig
+  - hiding in the shadows
+  style: A hyper-realistic digital painting of
+  subject: A leprechaun with a mischievous grin, emerald-green coat, and a hat adorned
+    with a four-leaf clover
+- concept_token: fairy
+  settings:
+  - dressed in a cloak of spider silk
+  - tending to forest creatures
+  - in a mystical glen
+  - frolicking in the air
+  - wearing a garland of fireflies
+  - sleeping in a flower
+  style: A hyper-realistic digital painting of
+  subject: A fairy
+- concept_token: leprechaun
+  settings:
+  - in a clover field
+  - mending a shoe
+  - inside a hollow oak tree
+  - sharing a pint
+  - at a village fair
+  - hiding a pot of gold
+  style: ''
+  subject: A leprechaun
+foods:
+- concept_token: loaf of bread
+  settings:
+  - cooling on a rustic wooden board
+  - sliced and served with butter
+  - displayed at a bustling farmer's market
+  - resting in a woven basket
+  - paired with a bowl of hearty soup
+  style: A warm countryside painting of
+  subject: A crusty loaf of freshly baked bread
+- concept_token: spaghetti
+  settings:
+  - topped with a rich tomato sauce
+  - sprinkled with grated Parmesan cheese
+  - served with garlic bread on the side
+  - twirled on a fork over a plate
+  - accompanied by a glass of red wine
+  style: An Italian-inspired artwork of
+  subject: A steaming plate of spaghetti with marinara sauce
+- concept_token: croissant
+  settings:
+  - displayed in a Parisian bakery
+  - paired with a cup of espresso
+  - placed on a white porcelain plate
+  - in a basket lined with cloth
+  - "enjoyed at a sunny caf\xE9 table"
+  style: A soft and elegant painting of
+  subject: A buttery, flaky croissant
+- concept_token: salad
+  settings:
+  - tossed with fresh greens and vegetables
+  - topped with a tangy vinaigrette
+  - served in a wooden bowl
+  - garnished with sliced avocado
+  - paired with a crusty bread roll
+  style: A fresh and vibrant painting of
+  subject: A garden-fresh salad full of color
+- concept_token: macaron
+  settings:
+  - arranged delicately on a porcelain plate
+  - displayed in a Parisian patisserie
+  - served with a cup of tea
+  - stacked in a rainbow of colors
+  - wrapped in a gift box with ribbons
+  style: A refined and pastel-hued painting of
+  subject: A colorful array of delicate macarons
+- concept_token: hot chocolate
+  settings:
+  - topped with whipped cream and cocoa
+  - served in a cozy mug by the fire
+  - accompanied by marshmallows on the side
+  - in a festive holiday-themed cup
+  - placed on a wooden tray with cookies
+  style: A warm and comforting painting of
+  subject: A steaming cup of hot chocolate
+- concept_token: fried rice
+  settings:
+  - served in a traditional wok
+  - topped with green onions and egg
+  - accompanied by soy sauce on the side
+  - enjoyed at a bustling street market
+  - paired with a cup of jasmine tea
+  style: A vibrant Asian-inspired painting of
+  subject: A plate of colorful fried rice
+- concept_token: cupcake
+  settings:
+  - topped with swirls of buttercream
+  - sprinkled with edible glitter
+  - placed on a decorative stand
+  - served at a birthday celebration
+  - paired with a glass of milk
+  style: A fun and festive painting of
+  subject: A perfectly decorated cupcake
+- concept_token: bagel
+  settings:
+  - topped with cream cheese and lox
+  - "displayed in a cozy caf\xE9"
+  - served with a side of fresh fruit
+  - toasted to a golden brown
+  - wrapped for a quick breakfast
+  style: A simple and inviting painting of
+  subject: A classic bagel with delicious toppings
+- concept_token: chocolate cake
+  settings:
+  - displayed in a high-end bakery window
+  - served at a birthday celebration
+  - enjoyed with a glass of red wine
+  - highlighted in a dessert cookbook
+  - paired with a scoop of ice cream
+  - presented at a wedding reception
+  - enjoyed during a romantic dinner
+  style: A digital illustration of
+  subject: A decadent chocolate cake with layers of rich ganache and fresh strawberries
+- concept_token: bowl of ramen
+  settings:
+  - served in a bustling ramen shop
+  - highlighted in a food blog
+  - enjoyed on a chilly evening
+  - paired with a side of gyoza
+  - featured in a travel documentary
+  - prepared by a master chef
+  - showcased at a food festival
+  - enjoyed at a street food stall
+  style: A ukiyo-e style woodblock print of
+  subject: A steaming bowl of ramen with tender pork slices, soft-boiled egg, and
+    vibrant vegetables
+- concept_token: plate of pasta
+  settings:
+  - served at an Italian trattoria
+  - enjoyed during a family dinner
+  - highlighted in a cooking show
+  - paired with a glass of red wine
+  - featured in a gourmet magazine
+  - prepared by a nonna
+  - showcased at a food festival
+  - enjoyed in a rustic kitchen
+  style: A realistic still life painting of
+  subject: A plate of pasta with homemade marinara sauce, fresh basil, and parmesan
+    cheese
+- concept_token: ice cream sundae
+  settings:
+  - served at a vintage ice cream parlor
+  - enjoyed during a summer fair
+  - highlighted in a dessert cookbook
+  - paired with a brownie
+  - featured in a children's party
+  - prepared for a special treat
+  - displayed in a colorful bowl
+  style: A whimsical illustration of
+  subject: An indulgent ice cream sundae with chocolate syrup, whipped cream, and
+    a cherry on top
+- concept_token: Caesar salad
+  settings:
+  - served at an upscale bistro
+  - enjoyed during a light lunch
+  - highlighted in a health food magazine
+  - paired with grilled chicken
+  - featured in a cooking class
+  - prepared for a summer picnic
+  style: A botanical illustration of
+  subject: A classic Caesar salad with crisp romaine, parmesan, and creamy dressing
+- concept_token: bowl of gazpacho
+  settings:
+  - served at a Spanish tapas bar
+  - enjoyed during a summer evening
+  - highlighted in a culinary travel show
+  - paired with crusty bread
+  - featured in a Mediterranean cookbook
+  - prepared by a home chef
+  - showcased at a garden party
+  - enjoyed on a sunny patio
+  style: A cubist painting of
+  subject: A refreshing bowl of gazpacho with fresh tomatoes, cucumbers, and a drizzle
+    of olive oil
+- concept_token: cheesecake
+  settings:
+  - served at a New York deli
+  - enjoyed during a special celebration
+  - highlighted in a dessert cookbook
+  - paired with a glass of dessert wine
+  - featured in a pastry shop
+  - prepared for a holiday gathering
+  style: A detailed ink drawing of
+  subject: A creamy cheesecake with a graham cracker crust and a raspberry swirl
+- concept_token: bowl of chili
+  settings:
+  - served at a chili cook-off
+  - enjoyed during a cold winter day
+  - highlighted in a comfort food magazine
+  - paired with cornbread
+  - featured in a family recipe book
+  - prepared for a game day party
+  - showcased at a food truck
+  style: A Western-style painting of
+  subject: A hearty bowl of chili with ground beef, beans, and a spicy tomato sauce
+- concept_token: lobster roll
+  settings:
+  - served at a seaside shack
+  - enjoyed during a summer vacation
+  - highlighted in a seafood magazine
+  - paired with a crisp white wine
+  - featured in a coastal restaurant
+  - prepared for a beach picnic
+  style: A nautical-themed illustration of
+  subject: A succulent lobster roll with tender lobster meat and a buttery bun
+- concept_token: bowl of poke
+  settings:
+  - served at a Hawaiian luau
+  - enjoyed during a beach day
+  - highlighted in a health food magazine
+  - paired with a tropical smoothie
+  - featured in a poke bar
+  - prepared for a summer party
+  - showcased at a food festival
+  style: A vibrant watercolor of
+  subject: A fresh bowl of poke with marinated fish, avocado, and rice
+- concept_token: bowl of pad thai
+  settings:
+  - served at a Thai street food stall
+  - enjoyed during a warm evening
+  - highlighted in a travel documentary
+  - paired with a cold beer
+  - featured in an Asian cookbook
+  - prepared by a local chef
+  - showcased at a food market
+  - enjoyed during a festival
+  style: A traditional Thai mural of
+  subject: A flavorful bowl of pad thai with shrimp, peanuts, and fresh lime
+humans:
+- concept_token: scientist
+  settings:
+  - conducting an experiment in a lab
+  - analyzing data on a computer
+  - presenting research at a conference
+  - observing through a microscope
+  - writing a research paper
+  style: A focused scene of
+  subject: A scientist with a lab coat
+- concept_token: lawyer
+  settings:
+  - defending a client in court
+  - negotiating a contract in an office
+  - reviewing legal documents
+  - advising clients on legal matters
+  - preparing for a trial
+  style: A professional scene of
+  subject: A lawyer in a suit
+- concept_token: waiter
+  settings:
+  - serving food to customers in a restaurant
+  - taking an order at a table
+  - cleaning a table after customers leave
+  - delivering drinks to a bar
+  - setting a table for dinner
+  style: A welcoming scene of
+  subject: A waiter in a uniform
+- concept_token: tailor
+  settings:
+  - measuring a client for a suit
+  - stitching a dress by hand
+  - fitting a shirt for alterations
+  - ironing clothes in a workshop
+  - arranging fabrics on a cutting table
+  style: A precise scene of
+  subject: A tailor with a sewing machine
+- concept_token: scientist
+  settings:
+  - conducting research in a lab
+  - analyzing samples under a microscope
+  - experimenting with chemicals
+  - presenting findings at a conference
+  - writing a research paper
+  style: A focused scene of
+  subject: A scientist with a lab coat
+- concept_token: barber
+  settings:
+  - cutting hair in a barbershop
+  - trimming a beard with scissors
+  - styling a client's hair
+  - shaving a client with a razor
+  - cleaning up the work area
+  style: A stylish scene of
+  subject: A barber with clippers
+- concept_token: dancer
+  settings:
+  - performing ballet on stage
+  - practicing in a dance studio
+  - rehearsing for a performance
+  - dancing at a wedding
+  - choreographing a new routine
+  style: A graceful scene of
+  subject: A dancer in a tutu
+- concept_token: dentist
+  settings:
+  - cleaning teeth in a dental office
+  - performing a root canal procedure
+  - explaining oral hygiene to a patient
+  - taking X-rays of teeth
+  - filling a cavity during an appointment
+  style: A medical scene of
+  subject: A dentist in scrubs
+- concept_token: lawyer
+  settings:
+  - cross-examining a witness in court
+  - preparing a legal case with colleagues
+  - advising clients on legal matters
+  - filing legal documents in an office
+  - negotiating a settlement agreement
+  style: A professional scene of
+  subject: A lawyer in a suit
+- concept_token: optometrist
+  settings:
+  - conducting an eye exam with a patient
+  - fitting eyeglasses for a client
+  - examining eye charts in a clinic
+  - discussing eye health with a patient
+  - testing vision with an automated device
+  style: A precise scene of
+  subject: An optometrist with a phoropter
+- concept_token: gardener
+  settings:
+  - planting flowers in a garden
+  - trimming hedges in a park
+  - watering plants in a greenhouse
+  - harvesting vegetables in a field
+  - arranging flowers in a basket
+  style: A vibrant scene of
+  subject: A gardener with gloves and a straw hat
+- concept_token: gentleman
+  settings:
+  - reading a newspaper
+  - at a vintage car show
+  - in a vineyard
+  - swinging a golf club
+  - listening to music
+  - dining at a fancy restaurant
+  style: A 3D animation of
+  subject: An elderly gentleman with distinguished gray hair, a neatly trimmed mustache,
+    and gentle blue eyes
+- concept_token: woman
+  settings:
+  - wearing a gardening apron
+  - at a seaside pier
+  - in a rose garden
+  - in a cozy library
+  - on a city street in the 1950s
+  style: A pixar style illustration of
+  subject: A woman with a slender figure, straight red hair, and freckles across the
+    nose
+- concept_token: woman
+  settings:
+  - baking cookies
+  - at a seaside pier
+  - in a rose garden
+  - dressed in a vintage 1950s outfit
+  - wearing a knitted shawl
+  - painting a landscape
+  - walking a dog
+  - "drinking coffee at a caf\xE9"
+  - taking photographs
+  - attending a fair
+  style: A hyper-realistic digital painting of
+  subject: A woman with a slender figure, straight red hair, and freckles across the
+    nose
+- concept_token: gentleman
+  settings:
+  - in a classic study
+  - at a vintage car show
+  - wearing a vineyard owner's attire
+  - on a golf course
+  - on a mountain
+  - walking with a cane
+  - discussing art at a gallery
+  - tending to a garden
+  - having a leisurely walk
+  - enjoying a sunset
+  style: A pixel art depiction of
+  subject: An elderly gentleman with distinguished gray hair, a neatly trimmed mustache,
+    and gentle blue eyes
+- concept_token: girl
+  settings:
+  - on a playground
+  - in a flower garden
+  - building a sandcastle
+  - reading a fairy tale book
+  - in a city park with autumn leaves
+  - drawing in a sketchbook
+  - taking a walk with friends
+  - shopping at a boutique
+  style: A hyper-realistic digital painting of
+  subject: a 16 years old girl with wavy chestnut hair, a slender frame, and soft
+    brown eyes
+- concept_token: boy
+  settings:
+  - green knitted hat
+  - at an urban skatepark
+  - in a dense forest
+  - dressed as an astronaut
+  - digging with a trowel
+  - climbing a tree
+  - attending a science fair
+  style: A hyper-realistic digital painting of
+  subject: A teenage boy with short, spiky black hair, a slight build, and dark brown
+    eyes
+- concept_token: child
+  settings:
+  - in a toy store
+  - exploring an exhibit
+  - in a backyard
+  - dressed in a prince costume
+  - in a treehouse
+  - playing with a puppy
+  - enjoying a carousel ride
+  - building a fort
+  style: A watercolor illustration of
+  subject: A male child with a round face, short ginger hair, and curious, wide eyes
+- concept_token: woman
+  settings:
+  - in a vintage kitchen
+  - dressed in a formal evening gown
+  - in a rose garden
+  - reading a novel
+  - shopping at a market
+  - visiting a museum
+  - having a picnic in a park
+  style: A photo of
+  subject: A woman with a slender figure, straight red hair, and freckles across the
+    nose
+- concept_token: woman
+  settings:
+  - baking a cake
+  - dressed in a formal evening gown
+  - wearing a chef's hat and apron
+  - holding a box
+  - in a snowy forest
+  - attending a holiday party
+  - singing at a concert
+  style: A watercolor illustration of
+  subject: A woman with a slender figure, straight red hair, and freckles across the
+    nose
+- concept_token: girl
+  settings:
+  - wearing a ballet dress
+  - in a flower garden
+  - building a sandcastle
+  - in a medieval castle courtyard
+  - wearing a birthday party hat
+  - dancing in a rainstorm
+  - writing in a journal
+  - practicing yoga
+  style: A pixel art depiction of
+  subject: a 16 years old girl with wavy chestnut hair, a slender frame, and soft
+    brown eyes
+inanimate:
+- concept_token: cloud
+  settings:
+  - floating on a postcard
+  - stuck on a book cover
+  - decorating a nursery wall
+  - on a birthday card
+  - placed on a phone case
+  style: A photo of
+  subject: A sticker of a fluffy white cloud
+- concept_token: starfish
+  settings:
+  - on a beach-themed photo album
+  - stuck on a notebook cover
+  - placed on a bathroom mirror
+  - decorating a summer tote bag
+  - on a seaside souvenir
+  style: A photo of
+  subject: A sticker of a detailed starfish
+- concept_token: butterfly
+  settings:
+  - placed on a floral scarf
+  - decorating a greeting card
+  - stuck on a wind chime
+  - on a flowerpot in the garden
+  - pinned on a corkboard
+  style: A photo of
+  subject: A sticker of a colorful butterfly
+- concept_token: rainbow
+  settings:
+  - decorating a birthday banner
+  - stuck on a car window
+  - placed on a child's lunchbox
+  - on a keychain for good luck
+  - on a window decal
+  style: A photo of
+  subject: A sticker of a vibrant rainbow
+- concept_token: umbrella
+  settings:
+  - decorating a rainy day card
+  - stuck on a luggage tag
+  - placed on a beach towel
+  - on a coffee cup sleeve
+  - as a window decal
+  style: A photo of
+  subject: A sticker of a colorful umbrella
+- concept_token: apple
+  settings:
+  - decorating a school notebook
+  - stuck on a kitchen fridge
+  - on a picnic basket lid
+  - on a lunchbox
+  - placed in a fruit bowl
+  style: A photo of
+  subject: A sticker of a shiny red apple
+- concept_token: bird
+  settings:
+  - perched on a window sill
+  - stuck on a travel mug
+  - decorating a greeting card
+  - placed on a beach towel
+  - on a decorative pillow
+  style: A photo of
+  subject: A sticker of a colorful bird
+- concept_token: pin
+  settings:
+  - decorating a corkboard
+  - on a jean jacket
+  - attached to a notebook cover
+  - on a travel map
+  - stuck on a greeting card
+  style: A photo of
+  subject: A sticker of a colorful pin
+- concept_token: hamburger
+  settings:
+  - placed on a restaurant menu
+  - decorating a food-themed poster
+  - stuck on a lunchbox
+  - on a picnic blanket
+  - in a fast food tray
+  style: A photo of
+  subject: A sticker of a delicious hamburger
+- concept_token: sun
+  settings:
+  - on a tropical vacation suitcase
+  - stuck on a lemonade stand
+  - decorating a summer beach ball
+  - on a travel-themed scrapbook
+  - placed on a sunlit balcony door
+  style: A photo of
+  subject: A sticker of a smiling yellow sun
+- concept_token: balloon
+  settings:
+  - on a birthday party invitation
+  - floating on a festival flyer
+  - stuck to a carnival ticket
+  - decorating a child's bedroom wall
+  - on a helium tank label
+  style: A photo of
+  subject: A sticker of a colorful balloon
+- concept_token: guitar
+  settings:
+  - on a musician's travel case
+  - stuck on a band poster
+  - decorating a music notebook
+  - placed on a record store bag
+  - on a concert ticket stub
+  style: A photo of
+  subject: A sticker of an acoustic guitar
+- concept_token: cupcake
+  settings:
+  - on a bakery box
+  - decorating a birthday invitation
+  - stuck on a sweet recipe book
+  - placed on a tea party favor
+  - on a dessert menu
+  style: A photo of
+  subject: A sticker of a frosted cupcake
+- concept_token: dog
+  settings:
+  - on a student's laptop
+  - amidst an action-packed skate session
+  - adding character to urban signage
+  - at a live music gig
+  - at an international airport
+  - in a bustling city center
+  - on a colorful street mural
+  style: A photo of
+  subject: A sticker of a cute corgi dog
+- concept_token: sofa
+  settings:
+  - occupied by a sleeping cat
+  - in a Victorian-style parlor
+  - in an outdoor patio
+  - in a bustling coffee shop
+  - with a fleece blanket
+  - in a modern living room
+  - beside a cozy fireplace
+  - in a chic urban loft
+  style: A hyper-realistic digital painting of
+  subject: A plush velvet sofa in a rich emerald green color with elegant, curved
+    armrests
+- concept_token: dog
+  settings:
+  - on a student's laptop
+  - on a skateboard in a park
+  - adding character to urban signage
+  - decorated with band logos
+  - on a traveler's suitcase
+  - in a crowded subway
+  - on a street artist's sketchbook
+  - at a bustling flea market
+  style: A hyper-realistic digital painting of
+  subject: A sticker of a cute corgi dog
+- concept_token: lamp
+  settings:
+  - with a classic green shade
+  - lighting up a late-night study session
+  - casting a warm glow in an elegant setting
+  - on a table
+  - on the outside
+  - in a minimalist workspace
+  - by a cozy reading nook
+  style: A 3D animation of
+  subject: A modern desk lamp with a metallic finish and an adjustable, minimalist
+    design
+- concept_token: mug
+  settings:
+  - being filled with steaming coffee
+  - sitting beside a laptop and notebooks
+  - on an office desk filled with papers
+  - on a potter's wheel being shaped
+  - on a balcony overlooking a cityscape
+  - in a cozy kitchen
+  - by a morning newspaper
+  style: A hyper-realistic digital painting of
+  subject: A red coffee mug shaped like a ball
+nature:
+- concept_token: sunflower
+  settings:
+  - stretching toward the midday sun
+  - in a farmer's field
+  - blooming in a garden patch
+  - surrounded by bees on a hot day
+  - against a bright blue sky
+  style: A vibrant oil painting of
+  subject: A cheerful sunflower
+- concept_token: deer
+  settings:
+  - grazing in a serene forest glade
+  - stepping lightly through the underbrush
+  - silhouetted against a rising moon
+  - resting in a grassy meadow
+  - bounding through a woodland clearing
+  style: A detailed pencil drawing of
+  subject: A graceful deer
+- concept_token: rainbow
+  settings:
+  - arching over a misty waterfall
+  - after a spring rainstorm
+  - across a golden wheat field
+  - over a peaceful valley
+  - against a backdrop of dark clouds
+  style: A vivid acrylic painting of
+  subject: A vibrant rainbow
+- concept_token: pine
+  settings:
+  - swaying in the mountain breeze
+  - standing tall in a snow-covered forest
+  - casting long shadows at dusk
+  - surrounded by mist in a quiet glade
+  - framing a scenic lakeside view
+  style: A serene pencil sketch of
+  subject: A tall pine tree
+- concept_token: butterfly
+  settings:
+  - flitting among garden flowers
+  - resting on a green leaf
+  - against a clear blue sky
+  - hovering over a brook
+  - dancing in tall grass
+  style: A watercolor illustration of
+  subject: A vibrant butterfly with iridescent wings
+- concept_token: oak
+  settings:
+  - tall in an autumn forest
+  - casting shadows at sunset
+  - blanketed in fresh snow
+  - in a field of wildflowers
+  - against rolling hills
+  style: A realist landscape painting of
+  subject: A mighty oak tree with spreading branches
+- concept_token: eagle
+  settings:
+  - soaring high above a cliff
+  - perched on a mountain peak
+  - gliding over a wide valley
+  - flying against a clear blue sky
+  - swooping over a forest
+  style: A majestic watercolor of
+  subject: A powerful eagle
+- concept_token: lily
+  settings:
+  - blooming in a tranquil pond
+  - placed in a ceramic vase
+  - floating on a still lake
+  - beside a peaceful riverbank
+  - surrounded by green leaves
+  style: A soft pastel illustration of
+  subject: A fragrant lily flower
+- concept_token: mushroom
+  settings:
+  - growing on the forest floor
+  - nestled among fallen leaves
+  - after a light spring rain
+  - near a mossy tree trunk
+  - surrounded by ferns and flowers
+  style: A delicate pen-and-ink drawing of
+  subject: A forest mushroom
+- concept_token: butterfly
+  settings:
+  - resting on a vibrant daisy
+  - fluttering around a lavender bush
+  - on the edge of a garden gate
+  - perched on a sunflower
+  - dancing in the spring breeze
+  style: A soft watercolor of
+  subject: A delicate butterfly
+- concept_token: fern
+  settings:
+  - unfurling in a forest glade
+  - growing at the base of a tree
+  - flourishing along a shaded trail
+  - carpeting a rocky cliffside
+  - swaying in the summer breeze
+  style: A naturalistic pencil drawing of
+  subject: A lush fern
+- concept_token: owl
+  settings:
+  - perched on a tree branch at dusk
+  - flying through a moonlit forest
+  - watching from a hollowed-out tree
+  - gliding over a misty field
+  - hooting in the silent night
+  style: A realistic watercolor of
+  subject: A wise owl in flight
+- concept_token: lily pad
+  settings:
+  - floating in a serene pond
+  - under a blooming lotus flower
+  - surrounded by dragonflies
+  - near the edge of a brook
+  - reflecting the moonlight
+  style: A tranquil painting of
+  subject: A green lily pad
+- concept_token: glacier
+  settings:
+  - towering above icy waters
+  - shimmering in the polar sun
+  - calving into the ocean
+  - surrounded by snowy peaks
+  - glowing in the arctic twilight
+  style: A dramatic oil painting of
+  subject: A massive glacier
+- concept_token: daisy
+  settings:
+  - blooming in a summer meadow
+  - in a vase on a windowsill
+  - swaying in a gentle breeze
+  - nestled in a grassy field
+  - with dew drops at dawn
+  style: A cheerful watercolor of
+  subject: A white daisy flower
+- concept_token: coral
+  settings:
+  - growing in a vibrant reef
+  - home to tiny fish
+  - illuminated by sunlight underwater
+  - swaying gently with the currents
+  - surrounded by blue water
+  style: A detailed underwater painting of
+  subject: A colorful coral
+- concept_token: willow
+  settings:
+  - by a peaceful riverbank
+  - swaying in a summer breeze
+  - shading a grassy picnic spot
+  - with branches touching the water
+  - against a cloudy sky
+  style: A serene landscape painting of
+  subject: A weeping willow tree
+- concept_token: dragonfly
+  settings:
+  - hovering above a pond
+  - resting on a reed
+  - darting through tall grass
+  - shimmering in the morning light
+  - near a field of wildflowers
+  style: A vibrant sketch of
+  subject: A delicate dragonfly
+- concept_token: coconut tree
+  settings:
+  - on a white sandy beach
+  - swaying in a tropical breeze
+  - shading a beachside hut
+  - with coconuts hanging high
+  - against a turquoise sky
+  style: A tropical watercolor of
+  subject: A tall coconut tree
+- concept_token: moss
+  settings:
+  - covering a forest stone
+  - growing on an old tree trunk
+  - carpeting a woodland floor
+  - near a babbling brook
+  - glistening after a rainstorm
+  style: A detailed pencil drawing of
+  subject: A patch of soft green moss
+- concept_token: sunflower field
+  settings:
+  - stretching to the horizon
+  - under a bright summer sky
+  - swaying in the warm wind
+  - alive with buzzing bees
+  - glowing in the evening light
+  style: A grand oil painting of
+  subject: A vibrant sunflower field
+- concept_token: tide pool
+  settings:
+  - teeming with colorful sea life
+  - reflecting the sky above
+  - nestled in rocky shoreline
+  - surrounded by barnacles and shells
+  - touched by the rising tide
+  style: A realistic painting of
+  subject: A lively tide pool
+- concept_token: waterfall mist
+  settings:
+  - rising from a powerful cascade
+  - drifting through a dense forest
+  - catching sunlight in a rainbow
+  - cooling a rocky riverbank
+  - surrounding a hidden grotto
+  style: A soft pastel painting of
+  subject: Mist from a waterfall
+- concept_token: aurora
+  settings:
+  - dancing across a polar sky
+  - reflected on an icy lake
+  - illuminating a snowy landscape
+  - against a mountain backdrop
+  - over a quiet arctic village
+  style: A celestial oil painting of
+  subject: A shimmering aurora borealis
+- concept_token: hedgehog
+  settings:
+  - sniffing among fallen leaves
+  - curled up in the grass
+  - exploring a forest trail
+  - hiding in a hollow log
+  - beside a blooming flower
+  style: A charming watercolor of
+  subject: A tiny hedgehog
+- concept_token: oak leaf
+  settings:
+  - turning golden in autumn
+  - floating down a gentle stream
+  - crunching underfoot on a trail
+  - resting on a park bench
+  - caught in a spider's web
+  style: A detailed pencil drawing of
+  subject: A textured oak leaf
+- concept_token: stag
+  settings:
+  - standing proud on a hillside
+  - silhouetted against the dawn
+  - walking through a frosty meadow
+  - under a canopy of stars
+  - by a peaceful forest stream
+  style: A majestic watercolor of
+  subject: A noble stag
+- concept_token: seagull
+  settings:
+  - soaring above crashing waves
+  - perched on a rocky outcrop
+  - circling a fishing boat
+  - calling out over the sea
+  - walking along a sandy shore
+  style: A lively watercolor of
+  subject: A seagull in flight
+- concept_token: lavender
+  settings:
+  - blooming in a sunny field
+  - filling a garden with fragrance
+  - tied in bundles on a table
+  - swaying in a gentle breeze
+  - with bees buzzing around
+  style: A calming watercolor of
+  subject: A cluster of lavender flowers
+- concept_token: raven
+  settings:
+  - perched on a barren branch
+  - cawing in the early morning mist
+  - gliding over a misty valley
+  - standing on a rocky outcrop
+  - silhouetted against the moon
+  style: A mysterious oil painting of
+  subject: A jet-black raven
+- concept_token: oak
+  settings:
+  - tall in an autumn forest
+  - casting shadows at sunset
+  - blanketed in fresh snow
+  - in a field of wildflowers
+  - against rolling hills
+  - shading a peaceful clearing
+  - in a lush green woodland
+  style: A realist landscape painting of
+  subject: A mighty oak tree with spreading branches
+- concept_token: butterfly
+  settings:
+  - flitting among garden flowers
+  - resting on a green leaf
+  - against a clear blue sky
+  - hovering over a brook
+  - dancing in tall grass
+  - with a rainbow backdrop
+  - landing on a child's hand
+  style: A watercolor illustration of
+  subject: A vibrant butterfly with iridescent wings
+- concept_token: snow leopard
+  settings:
+  - prowling in snowy mountains
+  - camouflaged on rocks
+  - stalking prey in the snow
+  - lying on a snowy ledge
+  - with wind ruffling its fur
+  - in winter morning light
+  - with jagged peaks behind
+  style: A photorealistic painting of
+  subject: A snow leopard with piercing blue eyes
+- concept_token: waterfall
+  settings:
+  - in a tropical forest
+  - under a sunny rainbow
+  - surrounded by mist
+  - sunlight through canopy
+  - in a mountain valley
+  - falling into a clear pool
+  - in morning light
+  style: A romantic landscape painting of
+  subject: A majestic waterfall cascading down rocky cliffs
+- concept_token: peacock
+  settings:
+  - strutting in a garden
+  - displaying feathers fully
+  - walking a garden path
+  - perched on a wall
+  - amid blooming flowers
+  - under a vine pergola
+  - on an estate lawn
+  style: A detailed Art Nouveau illustration of
+  subject: A magnificent peacock with iridescent feathers
+- concept_token: hummingbird
+  settings:
+  - in a flower garden
+  - amidst honeysuckles
+  - under a clear sky
+  - near a backyard feeder
+  - in a rainforest clearing
+  - with lush foliage behind
+  - during golden hour
+  style: A dynamic nature photograph of
+  subject: A tiny hummingbird hovering near a flower
+- concept_token: maple leaf
+  settings:
+  - falling from a tree
+  - on a leaf-covered path
+  - in a forest glade
+  - glowing in autumn light
+  - lying on a mossy rock
+  - under a clear sky
+  - drifting on a pond
+  style: A traditional Japanese ink wash painting of
+  subject: A bright red maple leaf in autumn
+- concept_token: dolphin
+  settings:
+  - in a tropical sea
+  - under midday sun
+  - near coral reefs
+  - beside a sunset boat
+  - amidst playful dolphins
+  - in open ocean waves
+  - near a small island
+  style: A vibrant marine watercolor of
+  subject: A playful dolphin leaping from the water
+- concept_token: bamboo
+  settings:
+  - in a Japanese garden
+  - swaying in the breeze
+  - against misty mountains
+  - sunlight through leaves
+  - in a bamboo forest
+  - beside a koi pond
+  - in morning fog
+  style: A traditional Chinese ink painting of
+  subject: A cluster of tall, graceful bamboo stalks
+- concept_token: polar bear
+  settings:
+  - in a glowing sunset
+  - on frozen tundra edge
+  - with a cub nearby
+  - among floating icebergs
+  - under twilight skies
+  - by snow-capped mountains
+  - in an Arctic storm
+  style: A serene Arctic oil painting of
+  subject: A majestic polar bear on a sheet of ice
+- concept_token: coral
+  settings:
+  - under tropical waters
+  - with darting fish
+  - among marine flora
+  - on sandy sea floor
+  - sunlight through water
+  - near an underwater cave
+  - swaying with seaweed
+  style: A bright and lively underwater painting of
+  subject: A vibrant coral reef teeming with life
+technoledge:
+- concept_token: robotic dog
+  settings:
+  - delivering items in a city
+  - guiding visually impaired people
+  - assisting in search-and-rescue missions
+  - playing fetch in a park
+  - patrolling a secure facility
+  style: A futuristic urban illustration of
+  subject: A robotic dog with sleek metallic limbs
+- concept_token: robotic horse
+  settings:
+  - running in a racing competition
+  - assisting farmers in fieldwork
+  - carrying heavy loads in rugged terrain
+  - patrolling a large estate
+  - providing transportation in rural areas
+  style: A dynamic outdoor illustration of
+  subject: A robotic horse with metallic limbs
+- concept_token: robotic rabbit
+  settings:
+  - delivering flowers in a botanical garden
+  - entertaining children in a playroom
+  - exploring urban parks with its owner
+  - assisting farmers in crop monitoring
+  - navigating offices delivering documents
+  style: A charming nature illustration of
+  subject: A robotic rabbit with fluffy ears
+- concept_token: robotic wolf
+  settings:
+  - guarding the perimeter of a secure facility
+  - tracking wildlife in a forest
+  - assisting in mountain rescues
+  - participating in a futuristic military exercise
+  - leading a pack of robots in exploration
+  style: A mysterious wilderness illustration of
+  subject: A robotic wolf with glowing eyes
+- concept_token: robotic dog
+  settings:
+  - delivering items in a city
+  - guiding visually impaired people
+  - assisting in search-and-rescue missions
+  - playing fetch in a park
+  - patrolling a secure facility
+  style: A futuristic urban illustration of
+  subject: A robotic dog
+- concept_token: robotic cheetah
+  settings:
+  - running in a race
+  - assisting in wildlife conservation
+  - tracking criminals in a city
+  - patrolling a security perimeter
+  - helping athletes train for speed
+  style: A high-speed action illustration of
+  subject: A robotic cheetah
+- concept_token: delivery robot
+  settings:
+  - transporting packages in a city
+  - running errands in a busy street
+  - delivering groceries to homes
+  - navigating a warehouse for parcels
+  - moving goods in a market
+  style: A compact delivery illustration of
+  subject: A delivery robot
+- concept_token: robotic elephant
+  settings:
+  - carrying heavy materials in construction
+  - assisting with tree planting
+  - patrolling wildlife reserves
+  - providing rides for tourists
+  - transporting goods over rough terrain
+  style: A majestic nature illustration of
+  subject: A robotic elephant
+- concept_token: robotic wolf
+  settings:
+  - patrolling secure areas
+  - assisting in wilderness exploration
+  - tracking wildlife in forests
+  - guarding perimeters in urban zones
+  - leading a robotic pack
+  style: A sleek nature illustration of
+  subject: A robotic wolf

resource/example.json

@@ -0,0 +1,160 @@
+{
+    "combinations": [
+        {
+            "id_prompt": "A hyper-realistic digital painting of a 16 years old girl.",
+            "frame_prompt_list": [
+                "in a flower garden",
+                "building a sandcastle",
+                "in a city park with autumn leaves"
+            ]
+        },
+        {
+            "id_prompt": "A vintage-style poster of a dog",
+            "frame_prompt_list": [
+                "playing a guitar at a country concert",
+                "sitting by a campfire under a starry sky",
+                "riding a skateboard through a bustling city",
+                "posing in front of a historical landmark",
+                "wearing an astronaut suit on the moon"
+            ]
+        },
+        {
+            "id_prompt": "A photo of a dog",
+            "frame_prompt_list": [
+                "dancing to music at a vibrant street festival",
+                "chasing a frisbee in a colorful park",
+                "wearing sunglasses while relaxing on a beach chair",
+                "posing for a photoshoot in a modern art gallery",
+                "jumping through a hoop at a circus performance",
+                "playing with a group of children at a playground",
+                "exploring a retro diner while wearing a bowtie"
+            ]
+        },
+        {
+            "id_prompt": "A mystical illustration of a wise wizard with a long, flowing beard",
+            "frame_prompt_list": [
+                "in a tower filled with ancient tomes and artifacts",
+                "casting a spell by the light of a full moon",
+                "standing before a magical portal in the forest",
+                "summoning a storm over a mountain peak",
+                "writing runes in a dusty spellbook",
+                "mixing potions in a dimly lit chamber",
+                "consulting a crystal ball"
+            ]
+        },
+        {
+            "id_prompt": "A pixar style illustration of a dragon",
+            "frame_prompt_list": [
+                "soaring gracefully through a rainbow sky",
+                "nestled among blooming cherry blossoms",
+                "playfully splashing in a sparkling lake"
+            ]
+        },
+        {
+            "id_prompt": "A whimsical painting of a delicate fairy",
+            "frame_prompt_list": [
+                "hovering over a moonlit pond",
+                "dancing on the petals of a giant flower",
+                "spreading fairy dust over a sleeping village",
+                "sitting on a mushroom in a magical forest",
+                "playing with fireflies at dusk"
+            ]
+        },
+        {
+            "id_prompt": "A hyper-realistic digital painting of an elderly gentleman",
+            "frame_prompt_list": [
+                "wearing a smoking jacket",
+                "at a vintage car show",
+                "wearing a vineyard owner's attire",
+                "on a golf course",
+                "at a classical music concert",
+                "painting a landscape"
+            ]
+        },
+        {
+            "id_prompt": "A vintage-style poster of a ceramic vase with an intricate floral pattern and a glossy, sky-blue glaze",
+            "frame_prompt_list": [
+                "holding a rare bouquet of flowers",
+                "displaying exotic orchids",
+                "complementing a corporate decor",
+                "containing delicate cherry blossoms",
+                "holding a vibrant arrangement of sunflowers",
+                "filled with a fresh bouquet of lavender and wild daisies"
+            ]
+        },
+        {
+            "id_prompt": "A photo of a happy hedgehog with its cheese",
+            "frame_prompt_list": [
+                "in an autumn forest",
+                "next to a tiny cheese wheel",
+                "sitting on a mushroom",
+                "under a picnic blanket",
+                "amid blooming spring flowers"
+            ]
+        },
+        {
+            "id_prompt": "A heartwarming illustration of a friendly troll",
+            "frame_prompt_list": [
+                "under a stone bridge covered in ivy",
+                "guarding a treasure chest in a dark cave",
+                "helping travelers across a river",
+                "sitting by a campfire in a foggy forest",
+                "building a shelter from fallen logs",
+                "fishing in a quiet stream at dusk",
+                "carving runes into a rock",
+                "resting under a large oak tree"
+            ]
+        },
+        {
+            "id_prompt": "A quaint illustration of a hobbit",
+            "frame_prompt_list": [
+                "in a cozy, round door cottage",
+                "sitting by a fireplace in a quaint home",
+                "working in a garden of vibrant vegetables",
+                "enjoying a feast under a starlit sky",
+                "reading a book in a sunlit meadow",
+                "walking through a peaceful village",
+                "celebrating with friends in a rustic tavern",
+                "exploring a hidden valley"
+            ]
+        },
+        {
+            "id_prompt": "A hyper-realistic digital painting of a young ginger boy with his ball",
+            "frame_prompt_list": [
+                "leaves scattering in a gentle breeze",
+                "standing in a quiet meadow",
+                "set against a vibrant sunset",
+                "in a busy street of people",
+                "by a colorful graffiti wall",
+                "amidst a field of blooming wildflowers"            ]
+        },
+        {
+            "id_prompt": "A cinematic portrait of a man and a woman standing together",
+            "frame_prompt_list": [
+                "under a sky full of stars",
+                "on a bustling city street at night",
+                "in a dimly lit jazz club",
+                "walking along a sandy beach at sunset",
+                "in a cozy coffee shop with large windows",
+                "in a vibrant art gallery surrounded by paintings",
+                "under an umbrella during a soft rain",
+                "on a quiet park bench amidst falling leaves",
+                "standing on a rooftop overlooking the city skyline"
+            ]
+        },
+        {
+            "id_prompt": "A cinematic portrait of a man, a woman, and a child",
+            "frame_prompt_list": [
+                "walking in a quiet park",
+                "under a starlit sky",
+                "by a rustic cabin",
+                "on a forest trail",
+                "by a peaceful lake",
+                "at a vibrant market",
+                "in a snowy street",
+                "by a carousel",
+                "on a picnic blanket"
+            ]
+        }
+    ]
+}

resource/gen_benchmark.py

@@ -0,0 +1,110 @@
+import argparse
+import os
+import yaml
+from main import generate_images, load_unet_controller
+from unet import utils
+import torch
+import queue 
+import threading  
+from tqdm import tqdm  # Import tqdm
+
+def main_ben(unet_controller, pipe, save_dir, id_prompt, frame_prompt_list, seed, window_length):
+    unet_controller.ipca_index = -1
+    unet_controller.ipca_time_step = -1
+    # Ensure each process uses its own assigned device
+    os.makedirs(save_dir, exist_ok=True)
+    images, story_image = generate_images(unet_controller, pipe, id_prompt, frame_prompt_list, save_dir, window_length, seed, verbose=False)
+    return images, story_image
+
+def process_instance(unet_controller, pipe, instance):
+    # Unpack instance and execute task
+    save_dir, id_prompt, frame_prompt_list, seed, window_length = instance
+    return main_ben(unet_controller, pipe, save_dir, id_prompt, frame_prompt_list, seed, window_length)
+
+def worker(device, unet_controller, pipe, task_queue, pbar):
+    # Process tasks until queue is empty
+    while not task_queue.empty():
+        instance = task_queue.get()
+        if instance is None:  # If None is encountered, stop the worker
+            break
+        # Process the instance
+        result = process_instance(unet_controller, pipe, instance)
+        # Log the completion
+        print(f"Finished processing {instance[1]}")  # Log the processed instance (id_prompt)
+        task_queue.task_done()  # Mark the task as done
+        pbar.update(1)  # Update the progress bar
+
+def main():
+    parser = argparse.ArgumentParser(description="Calculate image similarities using DreamSim or CLIP.")
+    parser.add_argument('--device', type=str, choices=['cuda:0', 'cuda:1', 'cuda'], default='cuda')
+    parser.add_argument('--save_dir', type=str,)    
+    parser.add_argument('--benchmark_path', type=str,)
+    parser.add_argument('--model_path', type=str, default='stabilityai/stable-diffusion-xl-base-1.0', help='Path to the model')
+    parser.add_argument('--precision', type=str, choices=["fp16", "fp32"], default="fp16", help='Model precision')
+    parser.add_argument('--window_length', type=int, default=10, help='Window length for story generation')
+    parser.add_argument('--num_gpus', type=int, default=2, help='Number of GPUs to use')
+    parser.add_argument('--fix_seed', type=int, default=42, help='-1 for random seed')
+    args = parser.parse_args()
+
+    # Create a list of devices
+    devices = [f'cuda:{i}' for i in range(args.num_gpus)]  # List of device names
+    if args.num_gpus == 1:
+        devices = [args.device]
+
+    # Load unet_controllers and pipes for each device
+    unet_controllers = {}
+    pipes = {}
+    for device in devices:
+        pipe, _ = utils.load_pipe_from_path(args.model_path, device, torch.float16 if args.precision == "fp16" else torch.float32, args.precision)
+        unet_controller = load_unet_controller(pipe, device)
+        unet_controller.Save_story_image = False
+        unet_controller.Prompt_embeds_mode = "svr-eot"
+        # unet_controller.Is_freeu_enabled = True
+        unet_controllers[device] = unet_controller
+        pipes[device] = pipe
+
+    # Load the benchmark data
+    with open(os.path.expanduser(args.benchmark_path), 'r') as file:
+        data = yaml.safe_load(file)
+
+    instances = []
+    for subject_domain, subject_domain_instances in data.items():
+        for index, instance in enumerate(subject_domain_instances):
+            id_prompt = f'{instance["style"]} {instance["subject"]}'
+            frame_prompt_list = instance["settings"]
+            save_dir = os.path.join(args.save_dir, f"{subject_domain}_{index}")
+            if args.fix_seed != -1:
+                seed = args.fix_seed
+            else:
+                import random
+                seed = random.randint(0, 2**32 - 1)
+            instances.append((save_dir, id_prompt, frame_prompt_list, seed, args.window_length))
+
+    # Create a task queue and populate it with instances
+    task_queue = queue.Queue()
+    for instance in instances:
+        task_queue.put(instance)
+
+    # Initialize tqdm progress bar
+    pbar = tqdm(total=len(instances))
+
+    # Create threads for each device to process instances
+    threads = []
+    for device in devices:
+        unet_controller = unet_controllers[device]
+        pipe = pipes[device]
+        thread = threading.Thread(target=worker, args=(device, unet_controller, pipe, task_queue, pbar))
+        threads.append(thread)
+        thread.start()
+        import time
+        time.sleep(1)  # Wait for 1 second before starting the next thread
+
+    # Wait for all threads to finish
+    for thread in threads:
+        thread.join()
+
+    # Close the progress bar
+    pbar.close()
+
+if __name__ == "__main__":
+    main()

unet/pipeline_stable_diffusion_xl.py

@@ -0,0 +1,1364 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+import torch.nn.functional as F
+from torch.autograd import grad
+
+import torch
+from transformers import (
+    CLIPImageProcessor,
+    CLIPTextModel,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import (
+    FromSingleFileMixin,
+    IPAdapterMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    TextualInversionLoaderMixin,
+)
+from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from diffusers.models.attention_processor import (
+    AttnProcessor2_0,
+    FusedAttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    is_invisible_watermark_available,
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+
+
+from unet.unet_controller import UNetController
+import unet.utils as utils 
+
+if is_invisible_watermark_available():
+    from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
+
+# if is_torch_xla_available():
+#     import torch_xla.core.xla_model as xm
+
+#     XLA_AVAILABLE = True
+# else:
+#     XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusionXLPipeline
+
+        >>> pipe = StableDiffusionXLPipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+        ... )
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "a photo of an astronaut riding a horse on mars"
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used,
+            `timesteps` must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class StableDiffusionXLPipeline(
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    FromSingleFileMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    TextualInversionLoaderMixin,
+    IPAdapterMixin,
+):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion XL.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder. Stable Diffusion XL uses the text portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        text_encoder_2 ([` CLIPTextModelWithProjection`]):
+            Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+            specifically the
+            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+            variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer_2 (`CLIPTokenizer`):
+            Second Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+            Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+            `stabilityai/stable-diffusion-xl-base-1-0`.
+        add_watermarker (`bool`, *optional*):
+            Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+            watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+            watermarker will be used.
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+    _optional_components = [
+        "tokenizer",
+        "tokenizer_2",
+        "text_encoder",
+        "text_encoder_2",
+        "image_encoder",
+        "feature_extractor",
+    ]
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+        "add_text_embeds",
+        "add_time_ids",
+        "negative_pooled_prompt_embeds",
+        "negative_add_time_ids",
+    ]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        image_encoder: CLIPVisionModelWithProjection = None,
+        feature_extractor: CLIPImageProcessor = None,
+        force_zeros_for_empty_prompt: bool = True,
+        add_watermarker: Optional[bool] = None,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+        self.default_sample_size = self.unet.config.sample_size
+
+        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+        if add_watermarker:
+            self.watermark = StableDiffusionXLWatermarker()
+        else:
+            self.watermark = None
+
+
+    def encode_prompt(
+        self,
+        prompt: str,
+        prompt_2: Optional[str] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+        unet_controller: Optional[UNetController] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder, lora_scale)
+
+            if self.text_encoder_2 is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            # textual inversion: process multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0] # [a, 1280]
+                
+                if unet_controller is not None and unet_controller.frame_prompt_express is not None:
+                    if unet_controller.Remove_pool_embeds:
+                        pooled_prompt_embeds = pooled_prompt_embeds.zero_()
+                        
+                    input_prompt_embeds = prompt_embeds.hidden_states[-2] if clip_skip is None else prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+                    alpha_weaken = unet_controller.Alpha_weaken 
+                    beta_weaken = unet_controller.Beta_weaken 
+                    alpha_strengthen = unet_controller.Alpha_enhance 
+                    beta_strengthen = unet_controller.Beta_enhance 
+                    frame_prompt_suppress = unet_controller.frame_prompt_suppress
+                    frame_prompt_express = unet_controller.frame_prompt_express
+                    
+                    if unet_controller.Prompt_embeds_mode == 'svr':
+                        for movement in frame_prompt_suppress:
+                            utils.swr_single_prompt_embeds(movement, input_prompt_embeds[0], prompt[0], unet_controller.tokenizer, alpha=alpha_weaken, beta=beta_weaken)
+                        utils.swr_single_prompt_embeds(frame_prompt_express, input_prompt_embeds[0], prompt[0], unet_controller.tokenizer, alpha=alpha_strengthen, beta=beta_strengthen)
+
+                    elif unet_controller.Prompt_embeds_mode == 'svr-eot':
+                        for movement in frame_prompt_suppress:
+                            utils.swr_single_prompt_embeds(movement, input_prompt_embeds[0], prompt[0], unet_controller.tokenizer, alpha=alpha_weaken, beta=beta_weaken, zero_eot=True)
+                        utils.swr_single_prompt_embeds(frame_prompt_express, input_prompt_embeds[0], prompt[0], unet_controller.tokenizer, alpha=alpha_strengthen, beta=beta_strengthen, zero_eot=True)
+
+                    elif unet_controller.Prompt_embeds_mode == "original":
+                        pass
+
+                    else:
+                        raise ValueError(f"Invalid prompt embeds mode: {unet_controller.Prompt_embeds_mode}")
+                    
+                    prompt_embeds = input_prompt_embeds
+
+                elif unet_controller is not None and unet_controller.frame_prompt_express_list is not None:
+                    
+                    if unet_controller.Remove_pool_embeds:
+                        pooled_prompt_embeds = pooled_prompt_embeds.zero_()
+                        
+                    input_prompt_embeds = prompt_embeds.hidden_states[-2] if clip_skip is None else prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+                    alpha_weaken = unet_controller.Alpha_weaken 
+                    beta_weaken = unet_controller.Beta_weaken 
+                    alpha_strengthen = unet_controller.Alpha_enhance 
+                    beta_strengthen = unet_controller.Beta_enhance 
+                    frame_prompt_suppress_list = unet_controller.frame_prompt_suppress_list
+                    frame_prompt_express_list = unet_controller.frame_prompt_express_list
+
+                    for index, (frame_prompt_suppress, frame_prompt_express) in enumerate(zip(frame_prompt_suppress_list, frame_prompt_express_list)):
+                    
+                        if unet_controller.Prompt_embeds_mode == 'svr':
+                            for movement in frame_prompt_suppress:
+                                utils.swr_single_prompt_embeds(movement, input_prompt_embeds[index], prompt[index], unet_controller.tokenizer, alpha=alpha_weaken, beta=beta_weaken)
+                            utils.swr_single_prompt_embeds(frame_prompt_express, input_prompt_embeds[index], prompt[index], unet_controller.tokenizer, alpha=alpha_strengthen, beta=beta_strengthen)
+
+                        elif unet_controller.Prompt_embeds_mode == "original":
+                            pass
+
+                        else:
+                            raise ValueError(f"Invalid prompt embeds mode: {unet_controller.Prompt_embeds_mode}")
+                        
+                    prompt_embeds = input_prompt_embeds
+
+                else: # original
+                    if clip_skip is None:
+                        prompt_embeds = prompt_embeds.hidden_states[-2]
+                    else:
+                        # "2" because SDXL always indexes from the penultimate layer.
+                        prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) # [a, 77, 2048]
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    negative_prompt,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        if self.text_encoder_2 is not None:
+            prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+        else:
+            prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            if self.text_encoder_2 is not None:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+            else:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        if do_classifier_free_guidance:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+                bs_embed * num_images_per_prompt, -1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+    def prepare_ip_adapter_image_embeds(
+        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+    ):
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
+
+            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+                raise ValueError(
+                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                )
+
+            image_embeds = []
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    single_ip_adapter_image, device, 1, output_hidden_state
+                )
+                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+                single_negative_image_embeds = torch.stack(
+                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
+                )
+
+                if do_classifier_free_guidance:
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                    single_image_embeds = single_image_embeds.to(device)
+
+                image_embeds.append(single_image_embeds)
+        else:
+            repeat_dims = [1]
+            image_embeds = []
+            for single_image_embeds in ip_adapter_image_embeds:
+                if do_classifier_free_guidance:
+                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                    single_negative_image_embeds = single_negative_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+                    )
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                else:
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                image_embeds.append(single_image_embeds)
+
+        return image_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        ip_adapter_image=None,
+        ip_adapter_image_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+            raise ValueError(
+                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+            )
+
+        if ip_adapter_image_embeds is not None:
+            if not isinstance(ip_adapter_image_embeds, list):
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+                )
+            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+                )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None, same=False):
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+            if same: # set all prompts' init_latent to the same
+                latents_pivot = latents[0]
+                for index in range(shape[0] - 1):
+                    latents[index+1] = latents_pivot
+        else:
+            latents = latents.to(device)
+        
+        # print(torch.mean(latents[0]))
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def _get_add_time_ids(
+        self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+    ):
+        add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        return add_time_ids
+
+    def upcast_vae(self):
+        dtype = self.vae.dtype
+        self.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+                FusedAttnProcessor2_0,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            self.vae.post_quant_conv.to(dtype)
+            self.vae.decoder.conv_in.to(dtype)
+            self.vae.decoder.mid_block.to(dtype)
+
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+    @property
+    def cross_attention_kwargs(self):
+        return self._cross_attention_kwargs
+
+    @property
+    def denoising_end(self):
+        return self._denoising_end
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 5.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        original_size: Optional[Tuple[int, int]] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Optional[Tuple[int, int]] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        unet_controller: Optional[UNetController] = None,
+        **kwargs,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+                Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+                Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+                if `do_classifier_free_guidance` is set to `True`.
+                If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+                Guidance rescale factor should fix overexposure when using zero terminal SNR.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+            [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+            )
+
+        # 0. Default height and width to unet
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+        self._interrupt = False
+
+        # 1.1 set unet_controller_parameter
+        if unet_controller is not None:
+                if prompt_2 is not None or negative_prompt_2 is not None:
+                    exit("current not support prompt_2 and negatvie_prompt_2")
+                unet_controller.do_classifier_free_guidance = self.do_classifier_free_guidance
+                if not isinstance(prompt, list):
+                    prompt_ = [prompt]
+                else:
+                    prompt_ = prompt
+                if not isinstance(negative_prompt, list):
+                    negative_prompt_ = [negative_prompt]
+                else:
+                    negative_prompt_ = negative_prompt
+                unet_controller.prompts = prompt_
+                unet_controller.negative_prompt = negative_prompt_
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+        )
+
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
+            unet_controller=unet_controller,
+        )
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+            same=(unet_controller.Use_same_latents if unet_controller is not None else False)
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Prepare added time ids & embeddings
+        add_text_embeds = pooled_prompt_embeds
+        if self.text_encoder_2 is None:
+            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+        else:
+            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+        add_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 8.1 Apply denoising_end
+        if (
+            self.denoising_end is not None
+            and isinstance(self.denoising_end, float)
+            and self.denoising_end > 0
+            and self.denoising_end < 1
+        ):
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+            timesteps = timesteps[:num_inference_steps]
+
+        # 9. Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+                
+                if unet_controller is not None:
+                    unet_controller.current_time_step = i
+                
+                # Expand the latents if we are doing classifier-free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # Predict the noise residual
+                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+                if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+                    added_cond_kwargs["image_embeds"] = image_embeds
+
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    unet_controller=unet_controller,
+                    return_dict=False,
+                )[0]   
+
+                # Perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+                # Compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+                    negative_pooled_prompt_embeds = callback_outputs.pop(
+                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+                    )
+                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+                    negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+                # if XLA_AVAILABLE:
+                #     xm.mark_step()
+
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            # apply watermark if available
+            if self.watermark is not None:
+                image = self.watermark.apply_watermark(image)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)

unet/unet.py

@@ -0,0 +1,599 @@
+# modified from the https://github.com/cloneofsimo/minSDXL
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.configuration_utils import ConfigMixin
+from typing import Optional
+
+from unet.unet_controller import UNetController
+import unet.utils as utils
+# SDXL
+
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int = 320):
+        super().__init__()
+        self.num_channels = num_channels
+
+    def forward(self, timesteps):
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(
+            half_dim, dtype=torch.float32, device=timesteps.device
+        )
+        exponent = exponent / (half_dim - 0.0)
+
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+
+        sin_emb = torch.sin(emb)
+        cos_emb = torch.cos(emb)
+        emb = torch.cat([cos_emb, sin_emb], dim=-1)
+
+        return emb
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_features, out_features):
+        super(TimestepEmbedding, self).__init__()
+        self.linear_1 = nn.Linear(in_features, out_features, bias=True)
+        self.act = nn.SiLU()
+        self.linear_2 = nn.Linear(out_features, out_features, bias=True)
+
+    def forward(self, sample):
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+
+        return sample
+
+
+class ResnetBlock2D(nn.Module):
+    def __init__(self, in_channels, out_channels, conv_shortcut=True):
+        super(ResnetBlock2D, self).__init__()
+        self.norm1 = nn.GroupNorm(32, in_channels, eps=1e-05, affine=True)
+        self.conv1 = nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        self.time_emb_proj = nn.Linear(1280, out_channels, bias=True)
+        self.norm2 = nn.GroupNorm(32, out_channels, eps=1e-05, affine=True)
+        self.dropout = nn.Dropout(p=0.0, inplace=False)
+        self.conv2 = nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        self.nonlinearity = nn.SiLU()
+        self.conv_shortcut = None
+        if conv_shortcut:
+            self.conv_shortcut = nn.Conv2d(
+                in_channels, out_channels, kernel_size=1, stride=1
+            )
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        temb = self.nonlinearity(temb)
+        temb = self.time_emb_proj(temb)[:, :, None, None]
+        hidden_states = hidden_states + temb
+        hidden_states = self.norm2(hidden_states)
+
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = input_tensor + hidden_states
+
+        return output_tensor
+
+
+class Attention(nn.Module):
+    def __init__(
+        self, inner_dim, cross_attention_dim=None, num_heads=None, dropout=0.0
+    ):
+        super(Attention, self).__init__()
+        if num_heads is None:
+            self.head_dim = 64
+            self.num_heads = inner_dim // self.head_dim
+        else:
+            self.num_heads = num_heads
+            self.head_dim = inner_dim // num_heads
+
+        self.scale = self.head_dim**-0.5
+        if cross_attention_dim is None:
+            cross_attention_dim = inner_dim
+        self.to_q = nn.Linear(inner_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(cross_attention_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(cross_attention_dim, inner_dim, bias=False)
+
+        self.to_out = nn.ModuleList(
+            [nn.Linear(inner_dim, inner_dim), nn.Dropout(dropout, inplace=False)]
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states=None, unet_controller: Optional[UNetController] = None):
+        q = self.to_q(hidden_states)
+        k = (
+            self.to_k(encoder_hidden_states)
+            if encoder_hidden_states is not None
+            else self.to_k(hidden_states)
+        )
+        v = (
+            self.to_v(encoder_hidden_states)
+            if encoder_hidden_states is not None
+            else self.to_v(hidden_states)
+        )
+        b, t, c = q.size()
+
+        q = q.view(q.size(0), q.size(1), self.num_heads, self.head_dim).transpose(1, 2)
+        k = k.view(k.size(0), k.size(1), self.num_heads, self.head_dim).transpose(1, 2)
+        v = v.view(v.size(0), v.size(1), self.num_heads, self.head_dim).transpose(1, 2)
+        
+
+        if (unet_controller is not None and unet_controller.Use_ipca and unet_controller.current_unet_position in unet_controller.Ipca_position
+              and encoder_hidden_states is not None and unet_controller.current_time_step >= unet_controller.Ipca_start_step):
+            
+            if unet_controller.do_classifier_free_guidance is True:
+                scores = torch.matmul(q, k.transpose(-2, -1)) * self.scale
+                attn_weights = torch.softmax(scores, dim=-1)  # this is only used by cross_attn_map store
+                ipca_attn_output = utils.ipca2(q,k,v,self.scale,unet_controller=unet_controller)
+                attn_output = ipca_attn_output
+            else:
+                exit("current doesn't support cfg=1.0")
+
+        else:
+            scores = torch.matmul(q, k.transpose(-2, -1)) * self.scale
+            attn_weights = torch.softmax(scores, dim=-1)
+            attn_output = torch.matmul(attn_weights, v)
+
+        attn_output = attn_output.transpose(1, 2).contiguous().view(b, t, c)
+
+        for layer in self.to_out:
+            attn_output = layer(attn_output)
+
+        return attn_output
+
+
+class GEGLU(nn.Module):
+    def __init__(self, in_features, out_features):
+        super(GEGLU, self).__init__()
+        self.proj = nn.Linear(in_features, out_features * 2, bias=True)
+
+    def forward(self, x):
+        x_proj = self.proj(x)
+        x1, x2 = x_proj.chunk(2, dim=-1)
+        return x1 * torch.nn.functional.gelu(x2)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, in_features, out_features):
+        super(FeedForward, self).__init__()
+
+        self.net = nn.ModuleList(
+            [
+                GEGLU(in_features, out_features * 4),
+                nn.Dropout(p=0.0, inplace=False),
+                nn.Linear(out_features * 4, out_features, bias=True),
+            ]
+        )
+
+    def forward(self, x):
+        for layer in self.net:
+            x = layer(x)
+        return x
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, hidden_size):
+        super(BasicTransformerBlock, self).__init__()
+        self.norm1 = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=True)
+        self.attn1 = Attention(hidden_size)
+        self.norm2 = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=True)
+        self.attn2 = Attention(hidden_size, 2048)
+        self.norm3 = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=True)
+        self.ff = FeedForward(hidden_size, hidden_size)
+
+    def forward(self, x, encoder_hidden_states=None, unet_controller: Optional[UNetController] = None):
+        residual = x
+
+        x = self.norm1(x)
+        x = self.attn1(x, unet_controller=unet_controller,)
+        x = x + residual
+
+        residual = x
+
+        x = self.norm2(x)
+        if encoder_hidden_states is not None:
+            x = self.attn2(x, encoder_hidden_states, unet_controller=unet_controller,)
+        else:
+            x = self.attn2(x, unet_controller=unet_controller,)
+        x = x + residual
+
+        residual = x
+
+        x = self.norm3(x)
+        x = self.ff(x)
+        x = x + residual
+        return x
+
+
+class Transformer2DModel(nn.Module):
+    def __init__(self, in_channels, out_channels, n_layers):
+        super(Transformer2DModel, self).__init__()
+        self.norm = nn.GroupNorm(32, in_channels, eps=1e-06, affine=True)
+        self.proj_in = nn.Linear(in_channels, out_channels, bias=True)
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(out_channels) for _ in range(n_layers)]
+        )
+        self.proj_out = nn.Linear(out_channels, out_channels, bias=True)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, unet_controller: Optional[UNetController] = None):
+        batch, _, height, width = hidden_states.shape
+        res = hidden_states
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
+            batch, height * width, inner_dim
+        )
+        hidden_states = self.proj_in(hidden_states)
+
+        for block in self.transformer_blocks:
+            hidden_states = block(hidden_states, encoder_hidden_states, unet_controller=unet_controller,)
+
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = (
+            hidden_states.reshape(batch, height, width, inner_dim)
+            .permute(0, 3, 1, 2)
+            .contiguous()
+        )
+
+        return hidden_states + res
+
+
+class Downsample2D(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(Downsample2D, self).__init__()
+        self.conv = nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=2, padding=1
+        )
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+class Upsample2D(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(Upsample2D, self).__init__()
+        self.conv = nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x):
+        x = F.interpolate(x, scale_factor=2.0, mode="nearest")
+        return self.conv(x)
+
+
+class DownBlock2D(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(DownBlock2D, self).__init__()
+        self.resnets = nn.ModuleList(
+            [
+                ResnetBlock2D(in_channels, out_channels, conv_shortcut=False),
+                ResnetBlock2D(out_channels, out_channels, conv_shortcut=False),
+            ]
+        )
+        self.downsamplers = nn.ModuleList([Downsample2D(out_channels, out_channels)])
+
+    def forward(self, hidden_states, temb):
+        output_states = []
+        for module in self.resnets:
+            hidden_states = module(hidden_states, temb)
+            output_states.append(hidden_states)
+
+        hidden_states = self.downsamplers[0](hidden_states)
+        output_states.append(hidden_states)
+
+        return hidden_states, output_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+    def __init__(self, in_channels, out_channels, n_layers, has_downsamplers=True):
+        super(CrossAttnDownBlock2D, self).__init__()
+        self.attentions = nn.ModuleList(
+            [
+                Transformer2DModel(out_channels, out_channels, n_layers),
+                Transformer2DModel(out_channels, out_channels, n_layers),
+            ]
+        )
+        self.resnets = nn.ModuleList(
+            [
+                ResnetBlock2D(in_channels, out_channels),
+                ResnetBlock2D(out_channels, out_channels, conv_shortcut=False),
+            ]
+        )
+        self.downsamplers = None
+        if has_downsamplers:
+            self.downsamplers = nn.ModuleList(
+                [Downsample2D(out_channels, out_channels)]
+            )
+
+    def forward(self, hidden_states, temb, encoder_hidden_states, unet_controller: Optional[UNetController] = None):
+        output_states = []
+        for resnet, attn in zip(self.resnets, self.attentions):
+            hidden_states = resnet(hidden_states, temb)
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                unet_controller=unet_controller,
+            )
+            output_states.append(hidden_states)
+
+        if self.downsamplers is not None:
+            hidden_states = self.downsamplers[0](hidden_states)
+            output_states.append(hidden_states)
+
+        return hidden_states, output_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+    def __init__(self, in_channels, out_channels, prev_output_channel, n_layers):
+        super(CrossAttnUpBlock2D, self).__init__()
+        self.attentions = nn.ModuleList(
+            [
+                Transformer2DModel(out_channels, out_channels, n_layers),
+                Transformer2DModel(out_channels, out_channels, n_layers),
+                Transformer2DModel(out_channels, out_channels, n_layers),
+            ]
+        )
+        self.resnets = nn.ModuleList(
+            [
+                ResnetBlock2D(prev_output_channel + out_channels, out_channels),
+                ResnetBlock2D(2 * out_channels, out_channels),
+                ResnetBlock2D(out_channels + in_channels, out_channels),
+            ]
+        )
+        self.upsamplers = nn.ModuleList([Upsample2D(out_channels, out_channels)])
+
+    def forward(
+        self, hidden_states, res_hidden_states_tuple, temb, encoder_hidden_states, unet_controller: Optional[UNetController] = None, 
+    ):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+            if unet_controller is not None and unet_controller.Is_freeu_enabled:
+                hidden_states, res_hidden_states = utils.apply_freeu(
+                    0 if unet_controller.current_unet_position == 'up0' else 1,
+                    hidden_states,
+                    res_hidden_states,
+                    s1=unet_controller.Freeu_parm['s1'],
+                    s2=unet_controller.Freeu_parm['s2'],
+                    b1=unet_controller.Freeu_parm['b1'],
+                    b2=unet_controller.Freeu_parm['b2'],
+                )
+
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+            hidden_states = resnet(hidden_states, temb)
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                unet_controller=unet_controller,
+            )
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states)
+
+        return hidden_states
+
+
+class UpBlock2D(nn.Module):
+    def __init__(self, in_channels, out_channels, prev_output_channel):
+        super(UpBlock2D, self).__init__()
+        self.resnets = nn.ModuleList(
+            [
+                ResnetBlock2D(out_channels + prev_output_channel, out_channels),
+                ResnetBlock2D(out_channels * 2, out_channels),
+                ResnetBlock2D(out_channels + in_channels, out_channels),
+            ]
+        )
+
+    def forward(self, hidden_states, res_hidden_states_tuple, temb=None):
+        for resnet in self.resnets:
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+    def __init__(self, in_features):
+        super(UNetMidBlock2DCrossAttn, self).__init__()
+        self.attentions = nn.ModuleList(
+            [Transformer2DModel(in_features, in_features, n_layers=10)]
+        )
+        self.resnets = nn.ModuleList(
+            [
+                ResnetBlock2D(in_features, in_features, conv_shortcut=False),
+                ResnetBlock2D(in_features, in_features, conv_shortcut=False),
+            ]
+        )
+
+    def forward(self, hidden_states, temb=None, encoder_hidden_states=None, unet_controller: Optional[UNetController] = None):
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                unet_controller=unet_controller,
+            )
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin):
+    def __init__(self):
+        super(UNet2DConditionModel, self).__init__() ## init child class first
+
+        # This is needed to imitate huggingface config behavior
+        # has nothing to do with the model itself
+        # remove this if you don't use diffuser's pipeline
+        # self.config = namedtuple(
+        #             "config", "in_channels addition_time_embed_dim sample_size time_cond_proj_dim"
+        #         )
+        # self.config.in_channels = 4
+        # self.config.addition_time_embed_dim = 256
+        # self.config.sample_size = 128
+        # self.config.time_cond_proj_dim = None
+
+        self.conv_in = nn.Conv2d(4, 320, kernel_size=3, stride=1, padding=1)
+        self.time_proj = Timesteps()
+        self.time_embedding = TimestepEmbedding(in_features=320, out_features=1280)
+        self.add_time_proj = Timesteps(256)
+        self.add_embedding = TimestepEmbedding(in_features=2816, out_features=1280)
+        self.down_blocks = nn.ModuleList(
+            [
+                DownBlock2D(in_channels=320, out_channels=320),
+                CrossAttnDownBlock2D(in_channels=320, out_channels=640, n_layers=2),
+                CrossAttnDownBlock2D(
+                    in_channels=640,
+                    out_channels=1280,
+                    n_layers=10,
+                    has_downsamplers=False,
+                ),
+            ]
+        )
+        self.up_blocks = nn.ModuleList(
+            [
+                CrossAttnUpBlock2D(
+                    in_channels=640,
+                    out_channels=1280,
+                    prev_output_channel=1280,
+                    n_layers=10,
+                ),
+                CrossAttnUpBlock2D(
+                    in_channels=320,
+                    out_channels=640,
+                    prev_output_channel=1280,
+                    n_layers=2,
+                ),
+                UpBlock2D(in_channels=320, out_channels=320, prev_output_channel=640),
+            ]
+        )
+        self.mid_block = UNetMidBlock2DCrossAttn(1280)
+        self.conv_norm_out = nn.GroupNorm(32, 320, eps=1e-05, affine=True)
+        self.conv_act = nn.SiLU()
+        self.conv_out = nn.Conv2d(320, 4, kernel_size=3, stride=1, padding=1)
+
+    
+    def forward(
+        self, sample, timesteps, encoder_hidden_states, added_cond_kwargs, unet_controller: Optional[UNetController] = None, **kwargs
+    ):
+        # Implement the forward pass through the model
+        timesteps = timesteps.expand(sample.shape[0])
+        t_emb = self.time_proj(timesteps).to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb)
+
+        text_embeds = added_cond_kwargs.get("text_embeds")
+        time_ids = added_cond_kwargs.get("time_ids")
+
+        time_embeds = self.add_time_proj(time_ids.flatten())
+        time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+        add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+        add_embeds = add_embeds.to(emb.dtype)
+        aug_emb = self.add_embedding(add_embeds)
+
+        emb = emb + aug_emb
+
+        sample = self.conv_in(sample)
+
+        # 3. down
+        if unet_controller is not None:
+            unet_controller.current_unet_position = 'down0'
+        
+        s0 = sample
+        sample, [s1, s2, s3] = self.down_blocks[0](
+            sample,
+            temb=emb,
+        )
+
+        if unet_controller is not None:
+            unet_controller.current_unet_position = 'down1'
+
+        # encoder_hidden_states is prompt_embedings, so here do cross_attn
+        sample, [s4, s5, s6] = self.down_blocks[1](
+            sample,
+            temb=emb, # time_embbeding
+            encoder_hidden_states=encoder_hidden_states, #[2,77,2048], 2 means two branch, 1 for prompt, 1 for negative prompt
+            unet_controller=unet_controller,
+        )
+
+        if unet_controller is not None:
+            unet_controller.current_unet_position = 'down2'
+
+        sample, [s7, s8] = self.down_blocks[2](
+            sample,
+            temb=emb,
+            encoder_hidden_states=encoder_hidden_states,
+            unet_controller=unet_controller,
+        )
+
+        # 4. mid
+        if unet_controller is not None:
+            unet_controller.current_unet_position = 'mid'
+
+        sample = self.mid_block(
+            sample, emb, encoder_hidden_states=encoder_hidden_states, unet_controller=unet_controller,
+        )
+
+        # 5. up
+        if unet_controller is not None:
+            unet_controller.current_unet_position = 'up0'
+
+        sample = self.up_blocks[0](
+            hidden_states=sample,
+            temb=emb,
+            res_hidden_states_tuple=[s6, s7, s8],
+            encoder_hidden_states=encoder_hidden_states,
+            unet_controller=unet_controller,
+        )
+
+        if unet_controller is not None:
+            unet_controller.current_unet_position = 'up1'
+
+        sample = self.up_blocks[1](
+            hidden_states=sample,
+            temb=emb,
+            res_hidden_states_tuple=[s3, s4, s5],
+            encoder_hidden_states=encoder_hidden_states,
+            unet_controller=unet_controller,
+        )
+
+        if unet_controller is not None:
+            unet_controller.current_unet_position = 'up2'
+
+        sample = self.up_blocks[2](
+            hidden_states=sample,
+            temb=emb,
+            res_hidden_states_tuple=[s0, s1, s2],
+        )
+
+        # 6. post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        return [sample]

unet/unet_controller.py

@@ -0,0 +1,73 @@
+import torch
+
+
+class UNetController():
+    # Static variables (Hyperparameters)
+    Is_freeu_enabled = False
+    Freeu_parm = {'s1': 0.6, 's2': 0.4, 'b1': 1.1, 'b2': 1.2}
+
+    # Ipca parameters
+    Use_ipca = True
+    Ipca_position = ['down0', 'down1', 'down2', 'mid', 'up0', 'up1', 'up2']
+    Ipca_start_step = 0
+    Ipca_dropout = 0.0
+    Use_embeds_mask = True
+
+    # SVR parameters
+    Alpha_weaken = 0.01  # 0.01~0.5
+    Beta_weaken = 0.05  # 0.05~1.0
+    Alpha_enhance = -0.01  # -0.001~-0.02
+    Beta_enhance = 1.0  # 1.0~2.0
+
+    # SVR settings
+    Prompt_embeds_mode = 'svr'
+    Remove_pool_embeds = False
+    Prompt_embeds_start_step = 0
+
+    Store_qkv = True
+
+    # other settings
+    Use_same_latents = True
+    Use_same_init_noise = True
+    Save_story_image = True
+
+    def __init__(self):
+        self._variables = {}
+
+        ## Variables (updated during inference) ##
+        self.device = "cuda"
+        self.current_unet_position = 'down'  # down, mid or up
+        self.torch_dtype = torch.float16
+
+        self.prompts = None
+        self.negative_prompt = None
+        self.id_prompt = None
+        self.frame_prompt_express = None
+        self.frame_prompt_suppress = None
+
+        self.frame_prompt_express_list = None
+        self.frame_prompt_suppress_list = None
+
+        self.tokenizer = None
+        self.result_save_dir = None
+        self.current_time_step = None
+        self.do_classifier_free_guidance = None
+
+        self.q_store = {}
+        self.k_store = {}
+        self.v_store = {}
+
+        self.do_classifier_free_guidance = None
+        self.current_unet_position = None
+
+        self.ipca2_index = -1
+        self.ipca_time_step = -1
+        ## Variables End ##
+    
+
+    def print_attributes(self):
+        """
+        Prints all attributes and their values of the object.
+        """
+        for attr, value in vars(self).items():
+            print(f"{attr}: {value}")

unet/utils.py

@@ -0,0 +1,483 @@
+import torch
+from typing import Optional
+from PIL import Image
+from diffusers import AutoencoderKL, EulerDiscreteScheduler, EDMDPMSolverMultistepScheduler
+from transformers import (
+    CLIPTextModel,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+)
+from scipy.spatial.distance import cdist
+import numpy as np
+import unet.pipeline_stable_diffusion_xl as pipeline_stable_diffusion_xl
+from torch.fft import fftn, fftshift, ifftn, ifftshift
+from typing import Optional, Tuple
+
+from unet.unet import UNet2DConditionModel
+from unet.unet_controller import UNetController
+
+
+def ipca(q, k, v, scale, unet_controller: Optional[UNetController] = None): # eg. q: [4,20,1024,64] k,v: [4,20,77,64] 
+    q_neg, q_pos = torch.split(q, q.size(0) // 2, dim=0)
+    k_neg, k_pos = torch.split(k, k.size(0) // 2, dim=0)
+    v_neg, v_pos = torch.split(v, v.size(0) // 2, dim=0)
+
+    # 1. negative_attn
+
+    scores_neg = torch.matmul(q_neg, k_neg.transpose(-2, -1)) * scale
+    attn_weights_neg = torch.softmax(scores_neg, dim=-1)
+    attn_output_neg = torch.matmul(attn_weights_neg, v_neg)
+
+    # 2. positive_attn (we do ipca only on positive branch)
+
+    # 2.1 ipca 
+    k_plus = torch.cat(tuple(k_pos.transpose(-2, -1)), dim=2).unsqueeze(0).repeat(k_pos.size(0),1,1,1) # 𝐾+ = [𝐾1 ⊕ 𝐾2 ⊕ . . . ⊕ 𝐾𝑁 ]
+    v_plus = torch.cat(tuple(v_pos), dim=1).unsqueeze(0).repeat(v_pos.size(0),1,1,1) # 𝑉+ = [𝑉1 ⊕ 𝑉2 ⊕ . . . ⊕ 𝑉𝑁 ]
+
+
+    # 2.2 apply mask
+    if unet_controller is not None:
+        scores_pos = torch.matmul(q_pos, k_plus) * scale
+
+        
+        # 2.2.1 apply dropout mask
+        dropout_mask = gen_dropout_mask(scores_pos.shape, unet_controller, unet_controller.Ipca_dropout) # eg: [a,1024,154]   
+
+
+        # 2.2.3 apply embeds mask
+        if unet_controller.Use_embeds_mask:
+            apply_embeds_mask(unet_controller,dropout_mask, add_eot=False)
+
+        mask = dropout_mask
+
+        mask = mask.unsqueeze(1).repeat(1,scores_pos.size(1),1,1)
+        attn_weights_pos = torch.softmax(scores_pos + torch.log(mask), dim=-1)
+
+    else:
+        scores_pos = torch.matmul(q_pos, k_plus) * scale
+        attn_weights_pos = torch.softmax(scores_pos, dim=-1)
+
+
+    attn_output_pos = torch.matmul(attn_weights_pos, v_plus)
+    # 3. combine
+    attn_output = torch.cat((attn_output_neg, attn_output_pos), dim=0)
+
+    return attn_output
+
+
+def ipca2(q, k, v, scale, unet_controller: Optional[UNetController] = None): # eg. q: [4,20,1024,64] k,v: [4,20,77,64] 
+    if unet_controller.ipca_time_step != unet_controller.current_time_step:
+        unet_controller.ipca_time_step = unet_controller.current_time_step
+        unet_controller.ipca2_index = 0
+    else:
+        unet_controller.ipca2_index += 1
+
+    if unet_controller.Store_qkv is True:
+
+        key = f"cross {unet_controller.current_time_step} {unet_controller.current_unet_position} {unet_controller.ipca2_index}"
+        unet_controller.k_store[key] = k
+        unet_controller.v_store[key] = v
+
+        scores = torch.matmul(q, k.transpose(-2, -1)) * scale
+        attn_weights = torch.softmax(scores, dim=-1)
+        attn_output = torch.matmul(attn_weights, v)
+    else:
+        # batch > 1
+        if unet_controller.frame_prompt_express_list is not None:
+            batch_size = q.size(0) // 2
+            attn_output_list = []
+
+            for i in range(batch_size):
+                q_i = q[[i, i + batch_size], :, :, :]
+                k_i = k[[i, i + batch_size], :, :, :]
+                v_i = v[[i, i + batch_size], :, :, :]
+
+                q_neg_i, q_pos_i = torch.split(q_i, q_i.size(0) // 2, dim=0)
+                k_neg_i, k_pos_i = torch.split(k_i, k_i.size(0) // 2, dim=0)
+                v_neg_i, v_pos_i = torch.split(v_i, v_i.size(0) // 2, dim=0)
+
+                key = f"cross {unet_controller.current_time_step} {unet_controller.current_unet_position} {unet_controller.ipca2_index}"
+                q_store = q_i
+                k_store = unet_controller.k_store[key]
+                v_store = unet_controller.v_store[key]
+
+                q_store_neg, q_store_pos = torch.split(q_store, q_store.size(0) // 2, dim=0)
+                k_store_neg, k_store_pos = torch.split(k_store, k_store.size(0) // 2, dim=0)
+                v_store_neg, v_store_pos = torch.split(v_store, v_store.size(0) // 2, dim=0)    
+
+                q_neg = torch.cat((q_neg_i, q_store_neg), dim=0)
+                q_pos = torch.cat((q_pos_i, q_store_pos), dim=0)
+                k_neg = torch.cat((k_neg_i, k_store_neg), dim=0)
+                k_pos = torch.cat((k_pos_i, k_store_pos), dim=0)
+                v_neg = torch.cat((v_neg_i, v_store_neg), dim=0)
+                v_pos = torch.cat((v_pos_i, v_store_pos), dim=0)
+
+                q_i = torch.cat((q_neg, q_pos), dim=0)
+                k_i = torch.cat((k_neg, k_pos), dim=0)
+                v_i = torch.cat((v_neg, v_pos), dim=0)
+
+                attn_output_i = ipca(q_i, k_i, v_i, scale, unet_controller)
+                attn_output_i = attn_output_i[[0, 2], :, :, :]
+                attn_output_list.append(attn_output_i)
+            
+            attn_output_ = torch.cat(attn_output_list, dim=0)
+            attn_output = torch.zeros(size=(q.size(0), attn_output_i.size(1), attn_output_i.size(2), attn_output_i.size(3)), device=q.device, dtype=q.dtype)
+            for i in range(batch_size):
+                attn_output[i] = attn_output_[i*2]
+            for i in range(batch_size):
+                attn_output[i + batch_size] = attn_output_[i*2 + 1]
+        # batch = 1
+        else:
+            q_neg, q_pos = torch.split(q, q.size(0) // 2, dim=0)
+            k_neg, k_pos = torch.split(k, k.size(0) // 2, dim=0)
+            v_neg, v_pos = torch.split(v, v.size(0) // 2, dim=0)
+
+            key = f"cross {unet_controller.current_time_step} {unet_controller.current_unet_position} {unet_controller.ipca2_index}"
+            q_store = q
+            k_store = unet_controller.k_store[key]
+            v_store = unet_controller.v_store[key]
+
+            q_store_neg, q_store_pos = torch.split(q_store, q_store.size(0) // 2, dim=0)
+            k_store_neg, k_store_pos = torch.split(k_store, k_store.size(0) // 2, dim=0)
+            v_store_neg, v_store_pos = torch.split(v_store, v_store.size(0) // 2, dim=0)    
+
+            q_neg = torch.cat((q_neg, q_store_neg), dim=0)
+            q_pos = torch.cat((q_pos, q_store_pos), dim=0)
+            k_neg = torch.cat((k_neg, k_store_neg), dim=0)
+            k_pos = torch.cat((k_pos, k_store_pos), dim=0)
+            v_neg = torch.cat((v_neg, v_store_neg), dim=0)
+            v_pos = torch.cat((v_pos, v_store_pos), dim=0)
+
+            q = torch.cat((q_neg, q_pos), dim=0)
+            k = torch.cat((k_neg, k_pos), dim=0)
+            v = torch.cat((v_neg, v_pos), dim=0)
+
+            attn_output = ipca(q, k, v, scale, unet_controller)
+            attn_output = attn_output[[0, 2], :, :, :]
+    
+    return attn_output
+
+
+def apply_embeds_mask(unet_controller: Optional[UNetController],dropout_mask, add_eot=False):   
+    id_prompt = unet_controller.id_prompt
+    prompt_tokens = prompt2tokens(unet_controller.tokenizer,unet_controller.prompts[0])
+    
+    words_tokens = prompt2tokens(unet_controller.tokenizer,id_prompt)
+    words_tokens = [word for word in words_tokens if word != '<|endoftext|>' and word != '<|startoftext|>']
+    index_of_words = find_sublist_index(prompt_tokens,words_tokens)    
+    index_list = [index+77 for index in range(index_of_words, index_of_words+len(words_tokens))]
+    if add_eot:
+        index_list.extend([index+77 for index, word in enumerate(prompt_tokens) if word == '<|endoftext|>'])
+
+    mask_indices = torch.arange(dropout_mask.size(-1), device=dropout_mask.device)
+    mask = (mask_indices >= 78) & (~torch.isin(mask_indices, torch.tensor(index_list, device=dropout_mask.device)))
+    dropout_mask[0, :, mask] = 0
+
+
+def gen_dropout_mask(out_shape, unet_controller: Optional[UNetController], drop_out):
+    gen_length = out_shape[3]
+    attn_map_side_length = out_shape[2]
+
+    batch_num = out_shape[0]
+    mask_list = []
+    
+    for prompt_index in range(batch_num):
+        start = prompt_index * int(gen_length / batch_num)
+        end = (prompt_index + 1) * int(gen_length / batch_num)
+    
+        mask = torch.bernoulli(torch.full((attn_map_side_length,gen_length), 1 - drop_out, dtype=unet_controller.torch_dtype, device=unet_controller.device))        
+        mask[:, start:end] = 1
+
+        mask_list.append(mask)
+
+    concatenated_mask = torch.stack(mask_list, dim=0)
+    return concatenated_mask
+
+
+def load_pipe_from_path(model_path, device, torch_dtype, variant):
+    model_name = model_path.split('/')[-1]
+    if model_path.split('/')[-1] == 'playground-v2.5-1024px-aesthetic':
+        scheduler = EDMDPMSolverMultistepScheduler.from_pretrained(model_path, subfolder="scheduler", torch_dtype=torch_dtype, variant=variant,)
+    else:
+        scheduler = EulerDiscreteScheduler.from_pretrained(model_path, subfolder="scheduler", torch_dtype=torch_dtype, variant=variant,)
+    
+    if model_path.split('/')[-1] == 'Juggernaut-X-v10' or model_path.split('/')[-1] == 'Juggernaut-XI-v11':
+        variant = None
+
+    vae = AutoencoderKL.from_pretrained(model_path, subfolder="vae", torch_dtype=torch_dtype, variant=variant,)
+    tokenizer = CLIPTokenizer.from_pretrained(model_path, subfolder="tokenizer", torch_dtype=torch_dtype, variant=variant,)
+    tokenizer_2 = CLIPTokenizer.from_pretrained(model_path, subfolder="tokenizer_2", torch_dtype=torch_dtype, variant=variant,)
+    text_encoder = CLIPTextModel.from_pretrained(model_path, subfolder="text_encoder", torch_dtype=torch_dtype, variant=variant,)
+    text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(model_path, subfolder="text_encoder_2", torch_dtype=torch_dtype, variant=variant,)
+    unet_new = UNet2DConditionModel.from_pretrained(model_path, subfolder="unet", torch_dtype=torch_dtype, variant=variant,)
+    
+    pipe = pipeline_stable_diffusion_xl.StableDiffusionXLPipeline(
+        vae=vae,
+        text_encoder=text_encoder,
+        text_encoder_2=text_encoder_2,
+        tokenizer=tokenizer,
+        tokenizer_2=tokenizer_2,
+        unet=unet_new,
+        scheduler=scheduler,
+    )
+    pipe.to(device)
+
+    return pipe, model_name
+
+
+def get_max_window_length(unet_controller: Optional[UNetController],id_prompt, frame_prompt_list):
+    single_long_prompt = id_prompt
+    max_window_length = 0
+    for index, movement in enumerate(frame_prompt_list):
+        single_long_prompt += ' ' + movement
+        token_length = len(single_long_prompt.split())
+        if token_length >= 77:
+            break
+        max_window_length += 1
+    return max_window_length
+
+
+def movement_gen_story_slide_windows(id_prompt, frame_prompt_list, pipe, window_length, seed, unet_controller: Optional[UNetController], save_dir, verbose=True):  
+    import os
+    max_window_length = get_max_window_length(unet_controller,id_prompt,frame_prompt_list)
+    window_length = min(window_length,max_window_length)
+    if window_length < len(frame_prompt_list):
+        movement_lists = circular_sliding_windows(frame_prompt_list, window_length)
+    else:
+        movement_lists = [movement for movement in frame_prompt_list]
+    story_images = []
+
+
+    if verbose: 
+        print("seed:", seed)
+    generate = torch.Generator().manual_seed(seed)
+    unet_controller.id_prompt = id_prompt
+
+    for index, movement in enumerate(frame_prompt_list):
+        if unet_controller is not None:
+            if window_length < len(frame_prompt_list):
+                unet_controller.frame_prompt_suppress = movement_lists[index][1:]
+                unet_controller.frame_prompt_express = movement_lists[index][0]
+                gen_propmts = [f'{id_prompt} {" ".join(movement_lists[index])}']
+
+            else:
+                unet_controller.frame_prompt_suppress = movement_lists[:index] + movement_lists[index+1:]
+                unet_controller.frame_prompt_express = movement_lists[index]
+                gen_propmts = [f'{id_prompt} {" ".join(movement_lists)}']
+            
+            if verbose:
+                print(f"suppress: {unet_controller.frame_prompt_suppress}")
+                print(f"express: {unet_controller.frame_prompt_express}")
+                print(f'id_prompt: {id_prompt}')
+                print(f"gen_propmts: {gen_propmts}")
+
+
+        else:
+            gen_propmts = f'{id_prompt} {movement}'
+
+        if unet_controller is not None and unet_controller.Use_same_init_noise is True:     
+            generate = torch.Generator().manual_seed(seed)
+
+        images = pipe(gen_propmts, generator=generate, unet_controller=unet_controller).images
+        story_images.append(images[0])
+        images[0].save(os.path.join(save_dir, f'{id_prompt} {unet_controller.frame_prompt_express}.jpg'))
+
+
+    image_array_list = [np.array(pil_img) for pil_img in story_images]
+
+    # Concatenate images horizontally
+    story_image = np.concatenate(image_array_list, axis=1)
+    story_image = Image.fromarray(story_image.astype(np.uint8))
+
+    if unet_controller.Save_story_image:
+        story_image.save(os.path.join(save_dir, f'story_image_{id_prompt}.jpg'))
+
+    return story_images, story_image
+
+# this function set batch > 1 to generate multiple images at once
+def movement_gen_story_slide_windows_batch(id_prompt, frame_prompt_list, pipe, window_length, seed, unet_controller: Optional[UNetController], save_dir, batch_size=3):  
+    import os
+    max_window_length = get_max_window_length(unet_controller,id_prompt,frame_prompt_list)
+    window_length = min(window_length,max_window_length)
+    if window_length < len(frame_prompt_list):
+        movement_lists = circular_sliding_windows(frame_prompt_list, window_length)
+    else:
+        movement_lists = [movement for movement in frame_prompt_list]
+    story_images = []
+
+    print("seed:", seed)
+    generate = torch.Generator().manual_seed(seed)
+    unet_controller.id_prompt = id_prompt
+
+    gen_prompt_info_list = []
+    gen_prompt = None
+    for index, _ in enumerate(frame_prompt_list):
+        if window_length < len(frame_prompt_list):
+            frame_prompt_suppress = movement_lists[index][1:]
+            frame_prompt_express = movement_lists[index][0]
+            gen_prompt = f'{id_prompt} {" ".join(movement_lists[index])}'
+
+        else:
+            frame_prompt_suppress = movement_lists[:index] + movement_lists[index+1:]
+            frame_prompt_express = movement_lists[index]
+            gen_prompt = f'{id_prompt} {" ".join(movement_lists)}'
+
+        gen_prompt_info_list.append({'frame_prompt_suppress': frame_prompt_suppress, 'frame_prompt_express': frame_prompt_express})
+    
+    story_images = []
+    for i in range(0, len(gen_prompt_info_list), batch_size):
+        batch = gen_prompt_info_list[i:i + batch_size]
+        gen_prompts = [gen_prompt for _ in batch]
+        unet_controller.frame_prompt_express_list = [gen_prompt_info['frame_prompt_express'] for gen_prompt_info in batch]
+        unet_controller.frame_prompt_suppress_list = [gen_prompt_info['frame_prompt_suppress'] for gen_prompt_info in batch]
+                
+        if unet_controller is not None and unet_controller.Use_same_init_noise is True:     
+            generate = torch.Generator().manual_seed(seed)
+        
+        images = pipe(gen_prompts, generator=generate, unet_controller=unet_controller).images    
+        for index,image in enumerate(images):
+            story_images.append(image)
+            image.save(os.path.join(save_dir, f'{id_prompt} {unet_controller.frame_prompt_express_list[index]}.jpg'))
+
+    image_array_list = [np.array(pil_img) for pil_img in story_images]
+
+    # Concatenate images horizontally
+    story_image = np.concatenate(image_array_list, axis=1)
+    story_image = Image.fromarray(story_image.astype(np.uint8))
+
+    if unet_controller.Save_story_image:
+        story_image.save(os.path.join(save_dir, 'story_image.jpg'))
+
+    return story_images, story_image
+
+
+def prompt2tokens(tokenizer, prompt):
+    text_inputs = tokenizer(
+        prompt,
+        padding="max_length",
+        max_length=tokenizer.model_max_length,
+        truncation=True,
+        return_tensors="pt",
+    )
+    text_input_ids = text_inputs.input_ids
+    tokens = []
+    for text_input_id in text_input_ids[0]:
+        token = tokenizer.decoder[text_input_id.item()]
+        tokens.append(token)
+    return tokens
+
+
+def punish_wight(tensor, latent_size, alpha=1.0, beta=1.2, calc_similarity=False):
+    u, s, vh = torch.linalg.svd(tensor)
+    u = u[:,:latent_size]
+    zero_idx = int(latent_size * alpha)
+
+    if calc_similarity:
+        _s = s.clone()
+        _s *= torch.exp(-alpha*_s) * beta
+        _s[zero_idx:] = 0
+        _tensor = u @ torch.diag(_s) @ vh
+        dist = cdist(tensor[:,0].unsqueeze(0).cpu(), _tensor[:,0].unsqueeze(0).cpu(), metric='cosine')
+        print(f'The distance between the word embedding before and after the punishment: {dist}')
+    s *= torch.exp(-alpha*s) * beta
+    tensor = u @ torch.diag(s) @ vh
+    return tensor
+
+
+def swr_single_prompt_embeds(swr_words,prompt_embeds,prompt,tokenizer,alpha=1.0, beta=1.2, zero_eot=False):
+    punish_indices = []
+
+    prompt_tokens = prompt2tokens(tokenizer,prompt)
+    
+    words_tokens = prompt2tokens(tokenizer,swr_words)
+    words_tokens = [word for word in words_tokens if word != '<|endoftext|>' and word != '<|startoftext|>']
+    index_of_words = find_sublist_index(prompt_tokens,words_tokens)
+    
+    if index_of_words != -1:
+        punish_indices.extend([num for num in range(index_of_words, index_of_words+len(words_tokens))])
+    
+    if zero_eot:
+        eot_indices = [index for index, word in enumerate(prompt_tokens) if word == '<|endoftext|>']
+        prompt_embeds[eot_indices] *= 9e-1
+        pass
+    else:
+        punish_indices.extend([index for index, word in enumerate(prompt_tokens) if word == '<|endoftext|>'])
+
+    punish_indices = list(set(punish_indices))
+    
+    wo_batch = prompt_embeds[punish_indices]
+    wo_batch = punish_wight(wo_batch.T.to(float), wo_batch.size(0), alpha=alpha, beta=beta, calc_similarity=False).T.to(prompt_embeds.dtype)
+
+    prompt_embeds[punish_indices] = wo_batch
+
+
+def find_sublist_index(list1, list2):
+    for i in range(len(list1) - len(list2) + 1):
+        if list1[i:i + len(list2)] == list2:
+            return i
+    return -1  # If sublist is not found
+
+
+def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor":
+    """Fourier filter as introduced in FreeU (https://arxiv.org/abs/2309.11497).
+
+    This version of the method comes from here:
+    https://github.com/huggingface/diffusers/pull/5164#issuecomment-1732638706
+    """
+    x = x_in
+    B, C, H, W = x.shape
+
+    x = x.to(dtype=torch.float32)
+
+    # FFT
+    x_freq = fftn(x, dim=(-2, -1))
+    x_freq = fftshift(x_freq, dim=(-2, -1))
+
+    B, C, H, W = x_freq.shape
+    mask = torch.ones((B, C, H, W), device=x.device)
+
+    crow, ccol = H // 2, W // 2
+    mask[..., crow - threshold : crow + threshold, ccol - threshold : ccol + threshold] = scale
+    x_freq = x_freq * mask
+
+    # IFFT
+    x_freq = ifftshift(x_freq, dim=(-2, -1))
+    x_filtered = ifftn(x_freq, dim=(-2, -1)).real
+
+    return x_filtered.to(dtype=x_in.dtype)
+
+
+def apply_freeu(
+    resolution_idx: int, hidden_states: "torch.Tensor", res_hidden_states: "torch.Tensor", **freeu_kwargs
+) -> Tuple["torch.Tensor", "torch.Tensor"]:
+    """Applies the FreeU mechanism as introduced in https:
+    //arxiv.org/abs/2309.11497. Adapted from the official code repository: https://github.com/ChenyangSi/FreeU.
+
+    Args:
+        resolution_idx (`int`): Integer denoting the UNet block where FreeU is being applied.
+        hidden_states (`torch.Tensor`): Inputs to the underlying block.
+        res_hidden_states (`torch.Tensor`): Features from the skip block corresponding to the underlying block.
+        s1 (`float`): Scaling factor for stage 1 to attenuate the contributions of the skip features.
+        s2 (`float`): Scaling factor for stage 2 to attenuate the contributions of the skip features.
+        b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+        b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+    """
+    if resolution_idx == 0:
+        num_half_channels = hidden_states.shape[1] // 2
+        hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b1"]
+        res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s1"])
+    if resolution_idx == 1:
+        num_half_channels = hidden_states.shape[1] // 2
+        hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b2"]
+        res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s2"])
+
+    return hidden_states, res_hidden_states
+
+
+def circular_sliding_windows(lst, w):
+    n = len(lst)
+    windows = []
+    for i in range(n):
+        window = [lst[(i + j) % n] for j in range(w)]
+        windows.append(window)
+    return windows