# app.py

import gradio as gr
import torch
from diffusers import AutoPipelineForInpainting
from PIL import Image
import time

# --- Model Loading (Final, Most Stable Version) ---
print("Loading the definitive model for low-RAM CPU environment...")
# We are using the more modern and reliable SD 2.0 Inpainting model.
# This model is better packaged and less prone to loading errors.
model_id = "stabilityai/stable-diffusion-2-inpainting"

try:
    pipe = AutoPipelineForInpainting.from_pretrained(
        model_id,
        torch_dtype=torch.float32, # Use float32 for CPU compatibility
        safety_checker=None        # Proactively disable the safety checker to save memory
    )
    
    # Enable CPU offloading to prevent memory crashes. This is essential.
    pipe.enable_model_cpu_offload()
    
    print("Model loaded successfully. The application is ready.")

except Exception as e:
    print("="*80)
    print("A FATAL ERROR OCCURRED DURING MODEL LOADING. The app cannot start.")
    print(f"Error: {e}")
    print("This is likely due to the free hardware tier not having enough resources.")
    print("="*80)
    raise e


# --- Default "Magic" Prompts ---
DEFAULT_PROMPT = "photorealistic, 4k, ultra high quality, sharp focus, masterpiece, high detail, professional photo"
DEFAULT_NEGATIVE_PROMPT = "blurry, pixelated, distorted, deformed, ugly, disfigured, cartoon, anime, low quality, watermark, text"

# --- The Inpainting Function ---
# This function signature is correct for how Gradio's Image tool works.
def inpaint_image(image_and_mask, user_prompt, guidance_scale, num_steps, progress=gr.Progress(track_tqdm=True)):
    
    if image_and_mask is None or "image" not in image_and_mask or "mask" not in image_and_mask:
        raise gr.Error("Please upload an image and draw a mask on it first!")
    
    # The input is a dictionary with 'image' and 'mask' keys
    image = image_and_mask["image"].convert("RGB")
    mask = image_and_mask["mask"].convert("RGB")
    
    if user_prompt and user_prompt.strip():
        prompt = user_prompt
        negative_prompt = DEFAULT_NEGATIVE_PROMPT 
        print(f"Using custom prompt: '{prompt}'")
    else:
        prompt = DEFAULT_PROMPT
        negative_prompt = DEFAULT_NEGATIVE_PROMPT
        print(f"User prompt is empty. Using default 'General Fix' prompt.")
    
    print(f"Starting inpainting on CPU (with offloading)... This will be very slow.")
    start_time = time.time()
    
    result_image = pipe(
        prompt=prompt,
        image=image,
        mask_image=mask,
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=int(num_steps),
    ).images[0]
    
    end_time = time.time()
    print(f"Inpainting finished in {end_time - start_time:.2f} seconds.")
    return result_image


# --- Gradio User Interface ---
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🎨 AI Image Fixer (Definitive Version)")
    gr.Warning(
        "‼️ **PATIENCE REQUIRED!** This app is running on a free CPU. "
        "Generation will be **extremely slow (potentially 20-40 minutes)** due to memory-saving measures. "
        "This is necessary to prevent crashes. The progress bar will appear after you click the button."
    )
    
    with gr.Row():
        with gr.Column(scale=2):
            input_image = gr.Image(label="1. Upload & Mask Image", source="upload", tool="brush", type="pil")
            prompt_textbox = gr.Textbox(label="2. Describe Your Fix (Optional)", placeholder="Leave empty for a general fix")
            with gr.Accordion("Advanced Settings", open=False):
                guidance_scale = gr.Slider(minimum=0, maximum=20, value=8.0, label="Guidance Scale")
                num_steps = gr.Slider(minimum=10, maximum=50, step=1, value=20, label="Inference Steps (Fewer is faster)")
        with gr.Column(scale=1):
            output_image = gr.Image(label="Result", type="pil")

    submit_button = gr.Button("Fix It!", variant="primary")
    
    submit_button.click(
        fn=inpaint_image,
        inputs=[input_image, prompt_textbox, guidance_scale, num_steps],
        outputs=output_image
    )

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