SD3-ControlNet-Depth

Demo

import torch
from diffusers import StableDiffusion3ControlNetPipeline
from diffusers.models import SD3ControlNetModel, SD3MultiControlNetModel
from diffusers.utils import load_image

# load pipeline
controlnet = SD3ControlNetModel.from_pretrained("InstantX/SD3-Controlnet-Depth")
pipe = StableDiffusion3ControlNetPipeline.from_pretrained(
    "stabilityai/stable-diffusion-3-medium-diffusers",
    controlnet=controlnet
)
pipe.to("cuda", torch.float16)

# config
control_image = load_image("https://huggingface.co./InstantX/SD3-Controlnet-Depth/resolve/main/images/depth.jpeg")
prompt = "a panda cub, captured in a close-up, in forest, is perched on a tree trunk. good composition, Photography, the cub's ears, a fluffy black, are tucked behind its head, adding a touch of whimsy to its appearance. a lush tapestry of green leaves in the background. depth of field, National Geographic"
n_prompt = "bad hands, blurry, NSFW, nude, naked, porn, ugly, bad quality, worst quality"

# to reproduce result in our example
generator = torch.Generator(device="cpu").manual_seed(4000)
image = pipe(
    prompt, 
    negative_prompt=n_prompt, 
    control_image=control_image, 
    controlnet_conditioning_scale=0.5,
    guidance_scale=7.0,
    generator=generator
).images[0]
image.save('image.jpg')

Limitation

Due to the fact that only 1024*1024 pixel resolution was used during the training phase, the inference performs best at this size, with other sizes yielding suboptimal results.