app.py

import torch
import numpy as np
import gradio as gr
from segment_anything import sam_model_registry, SamAutomaticMaskGenerator
from tqdm import tqdm
from diffusers import StableDiffusionInpaintPipeline, DDIMScheduler
from PIL import Image
from huggingface_hub import hf_hub_download, login
import os
import time

device = "cuda" if torch.cuda.is_available() else "cpu"

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

model_path = hf_hub_download(
    repo_id="Vuvo11/segment_anything_model",
    filename="sam_vit_h_4b8939.pth",
    use_auth_token=hf_token
)
sam = sam_model_registry["vit_h"](checkpoint=model_path)
mask_generator = SamAutomaticMaskGenerator(sam)

scheduler = DDIMScheduler.from_pretrained("runwayml/stable-diffusion-inpainting", subfolder="scheduler")
pipe = StableDiffusionInpaintPipeline.from_pretrained(
    "runwayml/stable-diffusion-inpainting",
    scheduler=scheduler,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,  # 🔥 FP16 cho GPU
    cache_dir="./models",
    low_cpu_mem_usage=True
).to(device)

if torch.cuda.is_available():
    pipe.unet = torch.compile(pipe.unet)  # 🔥 Tối ưu tốc độ nếu chạy trên GPU

pipe.enable_attention_slicing()

def generate_mask(image):
    masks = mask_generator.generate(image)
    if len(masks) == 0:
        return np.zeros_like(image[:, :, 0])

    largest_mask = max(masks, key=lambda x: np.sum(x["segmentation"]))
    return (largest_mask["segmentation"] * 255).astype(np.uint8)

def inpaint(image, prompt, progress=gr.Progress()):
    progress(0, "Processing image...")

    mask = generate_mask(image)  
    progress(30, "Generating inpainting...")

    original_image = Image.fromarray(image).convert("RGB")
    mask_image = Image.fromarray(mask).convert("L")

    original_image = original_image.resize((384, 384))  # 🔥 Resize nhỏ hơn để xử lý nhanh hơn
    mask_image = mask_image.resize((384, 384))

    output = pipe(prompt=prompt, image=original_image, mask_image=mask_image, num_inference_steps=15).images[0]  # 🔥 Giảm số bước suy luận

    progress(100, "Completed!")
    return np.array(output)

with gr.Blocks() as interface:
    gr.Markdown("## 🎨 AI Furniture Inpainting (Optimized)")
    
    with gr.Row():
        image_input = gr.Image(type="numpy", label="Upload Image")
        prompt_input = gr.Textbox(label="Prompt (Describe what to add)")
    
    submit = gr.Button("Submit")
    output_image = gr.Image(label="Generated Image")

    submit.click(fn=inpaint, inputs=[image_input, prompt_input], outputs=output_image)

if __name__ == "__main__":
    interface.launch()