garment_pipeline.py

import torch
from PIL import Image
from diffusers.utils import load_image
from pipeline import FluxConditionalPipeline

def generate_with_garment(
    pipe,
    garment_image: Image.Image,
    text: str,
    gemini_prompt: bool = True,
    guidance: float = 3.5,
    i_guidance: float = 1.0,
    t_guidance: float = 1.0,
    device="cuda" if torch.cuda.is_available() else "cpu"
):
    """
    Generates an image of a model wearing the provided garment with a new background
    
    Args:
        pipe: The FluxConditionalPipeline instance
        garment_image: Image of the garment to keep in the generated output
        text: Text prompt describing the desired output (model, pose, background)
        gemini_prompt: Whether to enhance the prompt using Gemini
        guidance: General guidance scale
        i_guidance: Image-specific guidance scale
        t_guidance: Text-specific guidance scale
        device: The device to use for generation
        
    Returns:
        The generated image
    """
    # Process the garment image
    w, h, min_size = garment_image.size[0], garment_image.size[1], min(garment_image.size)
    garment_image = garment_image.crop(
        ((w - min_size) // 2, (h - min_size) // 2, (w + min_size) // 2, (h + min_size) // 2)
    ).resize((512, 512))
    
    # Prepare garment image as control image
    control_image = load_image(garment_image)
    
    # Enhance the prompt to focus on keeping the garment while changing the model and background
    enhanced_text = text
    if not "garment" in enhanced_text.lower() and not "clothing" in enhanced_text.lower():
        enhanced_text = f"A model wearing this garment, {text}"
    
    # Generate the image
    result_image = pipe(
        prompt=enhanced_text.strip(),
        negative_prompt="distorted garment, wrong clothing, deformed clothes",
        num_inference_steps=28,
        height=512,
        width=1024,
        guidance_scale=guidance,
        image=control_image,
        guidance_scale_real_i=i_guidance,  # Higher value to maintain garment fidelity
        guidance_scale_real_t=t_guidance,
        gemini_prompt=gemini_prompt,
    ).images[0]
    
    return result_image