app.py

from fastapi import FastAPI, File, UploadFile, Form
from fastapi.responses import FileResponse, JSONResponse
import uuid
import os
from PIL import Image
import torch
from diffusers import (
    StableDiffusionControlNetPipeline,
    ControlNetModel,
    UniPCMultistepScheduler,
    StableDiffusionInpaintPipeline,
    StableDiffusionPipeline,
)
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
from ip_adapter.ip_adapter import IPAdapter
import cv2
import numpy as np

app = FastAPI()

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32
SAVE_DIR = "/tmp/kitchen_ai"
os.makedirs(SAVE_DIR, exist_ok=True)

# Load ControlNet
print("⏳ Loading ControlNet Canny...")
try:
    controlnet = ControlNetModel.from_pretrained(
        "lllyasviel/sd-controlnet-canny", torch_dtype=dtype, cache_dir="/tmp/hf_models"
    )
    pipe_canny = StableDiffusionControlNetPipeline.from_pretrained(
        "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=dtype, cache_dir="/tmp/hf_models"
    ).to(device)
    pipe_canny.scheduler = UniPCMultistepScheduler.from_config(pipe_canny.scheduler.config)
    print("✅ ControlNet loaded.")
except Exception as e:
    pipe_canny = None
    print("❌ ControlNet failed:", e)

# Load Inpainting (ĐÃ FIX)
print("⏳ Loading Inpainting model...")
try:
    pipe_inpaint = StableDiffusionInpaintPipeline.from_pretrained(
        "stabilityai/stable-diffusion-2-inpainting",  # ✅ FIXED
        torch_dtype=dtype,
        cache_dir="/tmp/hf_models"
    ).to(device)
    print("✅ Inpainting model loaded.")
except Exception as e:
    pipe_inpaint = None
    print("❌ Inpainting load failed:", e)

# Load IP-Adapter
print("⏳ Loading IP-Adapter...")
try:
    base_pipe = StableDiffusionPipeline.from_pretrained(
        "runwayml/stable-diffusion-v1-5", torch_dtype=dtype, cache_dir="/tmp/hf_models"
    ).to(device)
    vision_model = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-large-patch14")
    image_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-large-patch14")
    ip_adapter = IPAdapter(base_pipe, vision_model, image_processor, ip_ckpt="ip-adapter_sd15.safetensors")
    print("✅ IP-Adapter loaded.")
except Exception as e:
    ip_adapter = None
    print("❌ IP-Adapter failed:", e)

# Helper
def prepare_canny(image_path):
    img = cv2.imread(image_path)
    img = cv2.resize(img, (768, 768))  # tăng từ 512 nếu muốn ảnh chi tiết hơn
    edge = cv2.Canny(img, 100, 200)
    edge = cv2.cvtColor(edge, cv2.COLOR_GRAY2RGB)
    return Image.fromarray(edge)

# Routes
@app.post("/transform/")
async def transform_image(prompt: str = Form(...), image: UploadFile = File(...)):
    if pipe_canny is None:
        return JSONResponse({"error": "Model not available"}, status_code=500)
    input_path = os.path.join(SAVE_DIR, f"input_{uuid.uuid4().hex}.png")
    output_path = os.path.join(SAVE_DIR, f"output_{uuid.uuid4().hex}.png")
    with open(input_path, "wb") as f:
        f.write(await image.read())
    control_image = prepare_canny(input_path)
    result = pipe_canny(prompt=prompt, image=control_image, num_inference_steps=25).images[0]
    result.save(output_path)
    os.remove(input_path)
    return JSONResponse({"image_url": f"/download/{os.path.basename(output_path)}"})

@app.post("/transform_inpaint/")
async def transform_inpaint(prompt: str = Form(...), image: UploadFile = File(...), mask: UploadFile = File(...)):
    if pipe_inpaint is None:
        return JSONResponse({"error": "Inpaint model not ready"}, status_code=500)
    input_path = os.path.join(SAVE_DIR, f"inpaint_input_{uuid.uuid4().hex}.png")
    mask_path = os.path.join(SAVE_DIR, f"mask_{uuid.uuid4().hex}.png")
    output_path = os.path.join(SAVE_DIR, f"inpaint_output_{uuid.uuid4().hex}.png")
    with open(input_path, "wb") as f:
        f.write(await image.read())
    with open(mask_path, "wb") as f:
        f.write(await mask.read())
    init_image = Image.open(input_path).convert("RGB").resize((512, 512))
    mask_image = Image.open(mask_path).convert("L").resize((512, 512))
    result = pipe_inpaint(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
    result.save(output_path)
    os.remove(input_path)
    os.remove(mask_path)
    return JSONResponse({"image_url": f"/download/{os.path.basename(output_path)}"})

@app.post("/transform_ref/")
async def transform_ref(prompt: str = Form(...), image: UploadFile = File(...), ref_image: UploadFile = File(...)):
    if ip_adapter is None:
        return JSONResponse({"error": "IP-Adapter not ready"}, status_code=500)
    input_path = os.path.join(SAVE_DIR, f"ref_input_{uuid.uuid4().hex}.png")
    ref_path = os.path.join(SAVE_DIR, f"ref_img_{uuid.uuid4().hex}.png")
    output_path = os.path.join(SAVE_DIR, f"ref_output_{uuid.uuid4().hex}.png")
    with open(input_path, "wb") as f:
        f.write(await image.read())
    with open(ref_path, "wb") as f:
        f.write(await ref_image.read())
    pil_image = Image.open(input_path).convert("RGB").resize((512, 512))
    ref_image_pil = Image.open(ref_path).convert("RGB").resize((224, 224))
    images = ip_adapter.generate(
        pil_image=pil_image,
        ref_image=ref_image_pil,
        prompt=prompt,
        scale=0.6,
        seed=42
    )
    images[0].save(output_path)
    return JSONResponse({
        "image_url": f"/download/{os.path.basename(output_path)}",
        "status": "success"
    })

@app.get("/download/{filename}")
async def get_image(filename: str):
    file_path = os.path.join(SAVE_DIR, filename)
    if not os.path.exists(file_path):
        return JSONResponse({"error": "File not found"}, status_code=404)
    return FileResponse(file_path, media_type="image/png", filename=filename)