app.py

import hashlib
import io
import torch
from pathlib import Path
from diffusers import ControlNetModel, StableDiffusionControlNetPipeline, UniPCMultistepScheduler
from PIL import Image, ImageOps
import gradio as gr

# ---- Model loading ----
CACHE_DIR = "./cache"
CNET_MODEL = "MrPio/Texture-Anything_CNet-SD15"
SD_MODEL = "stable-diffusion-v1-5/stable-diffusion-v1-5"

controlnet = ControlNetModel.from_pretrained(
    CNET_MODEL, cache_dir=CACHE_DIR, torch_dtype=torch.float16
)
pipe = StableDiffusionControlNetPipeline.from_pretrained(
    SD_MODEL,
    controlnet=controlnet,
    cache_dir=CACHE_DIR,
    torch_dtype=torch.float16,
    safety_checker=None,
)

# speed & memory optimizations
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
# pipe.enable_xformers_memory_efficient_attention()  # if xformers installed
# pipe.enable_model_cpu_offload()


def pil2hash(image: Image.Image) -> str:
    buffer = io.BytesIO()
    image.save(buffer, format="PNG")
    image_bytes = buffer.getvalue()
    return hashlib.sha256(image_bytes).hexdigest()


def caption2hash(caption: str) -> str:
    return hashlib.sha256(caption.encode()).hexdigest()


# ---- Inference function ----
def infer(caption: str, condition_image: Image.Image, steps: int = 20, seed: int = 0, invert: bool = False):
    print("Loading condition image")
    img = condition_image.convert("RGB")
    if invert:
        img = ImageOps.invert(img)
        print("Condition image inverted")
    cache_file = Path(f"inferences/{pil2hash(img)}_{caption2hash(caption)}.png")
    if cache_file.exists():
        return Image.open(cache_file)

    generator = torch.manual_seed(seed)
    print("Starting generation...")
    output = pipe(prompt=caption, image=img, num_inference_steps=steps, generator=generator).images[0]
    print("Caching result...")
    output.save(cache_file)
    return output


# ---- Gradio UI + API ----
with gr.Blocks() as demo:
    gr.Markdown("## ControlNet + Stable Diffusion 1.5")
    with gr.Row():
        txt = gr.Textbox(label="Prompt", placeholder="Describe the texture...")
        cond = gr.Image(type="pil", label="Condition Image")
    with gr.Row():
        steps = gr.Slider(1, 50, value=20, label="Inference Steps")
        seed = gr.Number(value=0, label="Seed (0 for random)")
        inv = gr.Checkbox(label="Invert UV colors?")
    btn = gr.Button("Generate")
    out = gr.Image(label="Output")

    btn.click(fn=infer, inputs=[txt, cond, steps, seed, inv], outputs=out)

# enable the standard gradio REST API (/run/predict)
demo.launch(server_name="0.0.0.0", server_port=7860, share=True)