app.py

from fastapi import FastAPI, UploadFile, File
from pydantic import BaseModel
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline
from huggingface_hub import hf_hub_download
from fastapi.responses import StreamingResponse
from PIL import Image
import torch
import io

app = FastAPI()

# =========================
# تحميل الموديلات من HuggingFace
# =========================
MODEL_PATH = hf_hub_download(
    repo_id="ebraam1/interior-sd-models",
    filename="Interior.safetensors"
)

LORA_PATH = hf_hub_download(
    repo_id="ebraam1/interior-sd-models",
    filename="Interior_lora.safetensors"
)

print("Loading base model...")

# =========================
# Load Stable Diffusion
# =========================
txt2img = StableDiffusionPipeline.from_single_file(
    MODEL_PATH,
    torch_dtype=torch.float16,
    safety_checker=None
).to("cpu")   # غيّرها لـ "cuda" لو GPU متاح

img2img = StableDiffusionImg2ImgPipeline.from_single_file(
    MODEL_PATH,
    torch_dtype=torch.float16,
    safety_checker=None
).to("cpu")

print("Loading LoRA...")

txt2img.load_lora_weights(LORA_PATH)
img2img.load_lora_weights(LORA_PATH)

print("LoRA loaded 🔥")

# =========================
# API Schema
# =========================
class Prompt(BaseModel):
    prompt: str

# =========================
# Helper: PIL → Bytes
# =========================
def to_bytes(img):
    buf = io.BytesIO()
    img.save(buf, format="PNG")
    buf.seek(0)
    return buf

# =========================
# TEXT → IMAGE
# =========================
@app.post("/txt2img")
def generate(data: Prompt):

    image = txt2img(
        data.prompt,
        cross_attention_kwargs={"scale": 0.8}
    ).images[0]

    return StreamingResponse(to_bytes(image), media_type="image/png")

# =========================
# IMAGE → IMAGE
# =========================
@app.post("/img2img")
async def img2img_api(file: UploadFile = File(...), prompt: str = ""):

    img = Image.open(io.BytesIO(await file.read())).convert("RGB").resize((512, 512))

    image = img2img(
        prompt=prompt,
        image=img,
        cross_attention_kwargs={"scale": 0.8}
    ).images[0]

    return StreamingResponse(to_bytes(image), media_type="image/png")