main.py

import io
from contextlib import asynccontextmanager
from fastapi import FastAPI, Request, HTTPException, Response
from fastapi.responses import StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
from fastapi.concurrency import run_in_threadpool
from rembg import remove, new_session
from PIL import Image

ml_models = {}

def warmup_model() -> bool:
    """Warm up the model with a small image to reduce first-request latency."""
    session = ml_models.get("isnet")
    if session is None:
        return False
    try:
        # Create a small test image
        img = Image.new('RGB', (64, 64), color=(255, 255, 255))
        buf = io.BytesIO()
        img.save(buf, format='PNG')
        remove(buf.getvalue(), session=session)
        return True
    except Exception as e:
        print(f"Warm-up failed: {e}")
        return False
        
@asynccontextmanager
async def lifespan(app: FastAPI):
    """
    Lifespan Context Manager:
    Loads the AI model ONLY when the server starts.
    Cleans it up when the server stops.
    """
    print("Server starting: Loading ISNet model into memory...")
    # Load the model once and store it in the dictionary
    ml_models["isnet"] = new_session("isnet-general-use")
    yield
    # Cleanup logic (runs on shutdown)
    ml_models.clear()
    print("Server shutting down: Model memory released.")

app = FastAPI(
    title="AI Background Removal Service",
    description="High-precision background removal using ISNet and Alpha Matting.",
    lifespan=lifespan
)



origins = [
    "http://127.0.0.1:5500",
    "http://localhost:5500",
    "https://toolboxesai.com"
]

app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

def process_image_sync(input_bytes: bytes, params: dict) -> bytes:
    """
    Synchronous wrapper for the heavy lifting.
    """
    # --- CHANGE START: Retrieve model from lifespan dictionary ---
    session = ml_models.get("isnet")
    if session is None:
        raise RuntimeError("Model not loaded")
    # --- CHANGE END ---

    # (Existing logic kept exactly as you requested)
    use_alpha_matting = params.get("alpha_matting", "false").lower() == "true"
    
    af_thresh = int(params.get("af", 240))
    ab_thresh = int(params.get("ab", 10))

    return remove(
        input_bytes,
        session=session, # Pass the retrieved session here
        alpha_matting=use_alpha_matting,
        alpha_matting_foreground_threshold=af_thresh,
        alpha_matting_background_threshold=ab_thresh
    )

@app.post(
    "/api/remove-background",
    summary="Remove Background with Advanced Controls",
    responses={
        200: {"content": {"image/png": {}, "image/webp": {}}},
        400: {"description": "Invalid input."},
        500: {"description": "Processing error."},
    }
)
async def http_remove_background(request: Request):
    """
    Removes background using ISNet model. 
    Supports 'alpha_matting=true' query param for high-quality edge preservation.
    """
    
    # 1. Validate Query Params
    query_params = request.query_params
    output_format = query_params.get("format", "png").lower()
    
    if output_format not in ["png", "webp"]:
        raise HTTPException(400, "Unsupported format. Use 'png' or 'webp'.")

    # 2. Validate Content-Type
    content_type = request.headers.get("content-type")
    if not content_type or not content_type.startswith("image/"):
        raise HTTPException(400, "Invalid Content-Type. Must be image/*")

    try:
        # 3. Read Input (Async I/O)
        input_bytes = await request.body()
        if not input_bytes:
            raise HTTPException(400, "No image data provided.")

        # 4. Process Image (CPU Bound - Offloaded to Threadpool)
        # This prevents the server from freezing while calculating alpha values.
        output_bytes = await run_in_threadpool(
            process_image_sync, 
            input_bytes, 
            query_params
        )
        
        # 5. Handle Format Conversion
        final_bytes = output_bytes
        media_type = "image/png"

        if output_format == "webp":
            # WebP conversion (also CPU bound, ideally offloaded, but fast enough here)
            try:
                pil_image = Image.open(io.BytesIO(output_bytes))
                webp_buffer = io.BytesIO()
                pil_image.save(webp_buffer, format="WEBP", quality=90)
                final_bytes = webp_buffer.getvalue()
                media_type = "image/webp"
            except Exception as e:
                # Fallback to PNG if WebP fails
                print(f"WebP conversion failed: {e}")
                final_bytes = output_bytes
                media_type = "image/png"
                output_format = "png"

        # 6. Return Stream
        return StreamingResponse(
            io.BytesIO(final_bytes), 
            media_type=media_type,
            headers={
                "Content-Disposition": f"attachment; filename=bg_removed.{output_format}"
            }
        )

    except Exception as e:
        print(f"Error: {e}")
        raise HTTPException(500, f"Processing failed: {str(e)}")

@app.get("/health")
async def health_check():
    """Health check with model warm-up - call this to keep model hot."""
    model_loaded = ml_models.get("isnet") is not None
    
    if model_loaded:
        warmup_success = await run_in_threadpool(warmup_model)
        return {
            "status": "healthy",
            "model_loaded": True,
            "model_warmed": warmup_success
        }
    
    return Response(
        content='{"status": "unhealthy", "model_loaded": false}',
        media_type="application/json",
        status_code=503
    )

@app.get("/ping")
def get_ping():
    return Response(content="pong", media_type="text/plain")
if __name__ == "__main__":
    import uvicorn
    uvicorn.run("main:app", host="0.0.0.0", port=7860, reload=True)