app.py

import io
import zipfile
from typing import Dict

import torch
from fastapi import FastAPI, File, UploadFile
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse, StreamingResponse
from PIL import Image
from rembg import remove
from diffusers import DiffusionPipeline


app = FastAPI(title="Zero123++ Inference API")

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

MODEL_ID = "sudo-ai/zero123plus-v1.2"
CUSTOM_PIPELINE = "sudo-ai/zero123plus-pipeline"

pipeline = None


def load_pipeline():
    global pipeline

    if pipeline is not None:
        return pipeline

    if not torch.cuda.is_available():
        raise RuntimeError(
            "CUDA GPU is not available. Please enable GPU hardware on the Hugging Face Space."
        )

    pipe = DiffusionPipeline.from_pretrained(
        MODEL_ID,
        custom_pipeline=CUSTOM_PIPELINE,
        torch_dtype=torch.float16,
        trust_remote_code=True,
    )

    pipe.to("cuda")
    pipe.enable_attention_slicing()

    pipeline = pipe
    return pipeline


def image_to_bytes(image: Image.Image, fmt: str = "PNG") -> bytes:
    buffer = io.BytesIO()
    image.save(buffer, format=fmt)
    buffer.seek(0)
    return buffer.getvalue()


def crop_selected_views(grid: Image.Image) -> Dict[str, Image.Image]:
    """
    Zero123++ output is expected as a 2-column x 3-row grid.
    We keep views 1, 3, 5, and 6 for LGM.
    """

    grid = grid.convert("RGB")
    w, h = grid.size

    cols, rows = 2, 3
    tile_w, tile_h = w // cols, h // rows

    selected = {
        1: "front_right",
        3: "right",
        5: "back_right",
        6: "back_left",
    }

    outputs = {}

    for idx_1based, name in selected.items():
        row = (idx_1based - 1) // cols
        col = (idx_1based - 1) % cols

        box = (
            col * tile_w,
            row * tile_h,
            (col + 1) * tile_w,
            (row + 1) * tile_h,
        )

        tile = grid.crop(box).resize((256, 256), Image.LANCZOS)

        # Remove background and paste on white
        tile_rgba = remove(tile.convert("RGBA"))
        white_bg = Image.new("RGBA", tile_rgba.size, (255, 255, 255, 255))
        white_bg.paste(tile_rgba, mask=tile_rgba.split()[3])

        outputs[f"view_{idx_1based}_{name}.png"] = white_bg.convert("RGB")

    return outputs


@app.get("/")
def root():
    return {
        "status": "running",
        "service": "zero123plus-inference",
        "model": MODEL_ID,
        "output": "6-view grid + cropped views 1, 3, 5, 6 for LGM",
    }


@app.get("/health")
def health():
    return {
        "status": "ok",
        "cuda_available": torch.cuda.is_available(),
        "cuda_device": torch.cuda.get_device_name(0)
        if torch.cuda.is_available()
        else None,
    }


@app.post("/generate")
async def generate(file: UploadFile = File(...), steps: int = 75):
    try:
        pipe = load_pipeline()

        contents = await file.read()
        input_image = Image.open(io.BytesIO(contents)).convert("RGB")

        with torch.inference_mode():
            result = pipe(input_image, num_inference_steps=steps).images[0]

        cropped_views = crop_selected_views(result)

        zip_buffer = io.BytesIO()

        with zipfile.ZipFile(zip_buffer, "w", zipfile.ZIP_DEFLATED) as zf:
            zf.writestr("multiview_grid.png", image_to_bytes(result))

            for filename, image in cropped_views.items():
                zf.writestr(filename, image_to_bytes(image))

        zip_buffer.seek(0)

        return StreamingResponse(
            zip_buffer,
            media_type="application/zip",
            headers={
                "Content-Disposition": "attachment; filename=zero123plus_outputs.zip"
            },
        )

    except Exception as e:
        return JSONResponse(
            status_code=500,
            content={
                "error": str(e),
                "message": "Zero123++ generation failed. Check Space logs for details.",
            },
        )