app.py

import gradio as gr
import torch
from PIL import Image
from models.transformer_sd3 import SD3Transformer2DModel
from pipeline_stable_diffusion_3_ipa import StableDiffusion3Pipeline
import os
from huggingface_hub import login

token = os.getenv("HF_TOKEN")
login(token=token)

# Model and paths
model_path = 'stabilityai/stable-diffusion-3.5-large'
ip_adapter_path = './ip-adapter.bin'
image_encoder_path = "google/siglip-so400m-patch14-384"
ref_img_path = './assets/1.jpg'  # Reference image path

# Load SD3.5 pipeline and components
transformer = SD3Transformer2DModel.from_pretrained(
    model_path, subfolder="transformer", torch_dtype=torch.bfloat16
)
pipe = StableDiffusion3Pipeline.from_pretrained(
    model_path, transformer=transformer, torch_dtype=torch.bfloat16
).to("cuda")

pipe.init_ipadapter(
    ip_adapter_path=ip_adapter_path,
    image_encoder_path=image_encoder_path,
    nb_token=64,
)


@gr.Interface()
def gui_generation(prompt: str, negative_prompt: str, ipadapter_scale: float, num_imgs: int):
    """
    Generate images based on prompt, negative prompt, and IP-Adapter scale.
    """
    ref_img = Image.open(ref_img_path).convert('RGB')  # Load reference image
    generator = torch.Generator("cuda").manual_seed(42)  # Reproducibility

    images = []
    for _ in range(num_imgs):
        output = pipe(
            width=1024,
            height=1024,
            prompt=prompt,
            negative_prompt=negative_prompt,
            num_inference_steps=24,
            guidance_scale=5.0,
            generator=generator,
            clip_image=ref_img,
            ipadapter_scale=ipadapter_scale,
        ).images[0]
        images.append(output)
    return images


# Gradio UI elements
prompt_box = gr.Textbox(label="Prompt", placeholder="Enter your generation prompt here")
negative_prompt_box = gr.Textbox(label="Negative Prompt", placeholder="e.g., lowres, worst quality")
ipadapter_slider = gr.Slider(0.1, 1.0, value=0.5, step=0.1, label="IP-Adapter Scale")
number_slider = gr.Slider(1, 5, value=1, step=1, label="Number of Images")
gallery = gr.Gallery(label="Generated Images", columns=[3], rows=[1], object_fit="contain", height="auto")

interface = gr.Interface(
    gui_generation,
    inputs=[prompt_box, negative_prompt_box, ipadapter_slider, number_slider],
    outputs=gallery,
    title="Stable Diffusion 3.5 Image Generation with IP-Adapter",
    description="Generate high-quality images with Stable Diffusion 3.5 Large and IP-Adapter guidance."
)

interface.launch()