utils.py

import torch
from PIL import Image
from diffusers import StableDiffusionInpaintPipeline
import numpy as np

pipe_cache = {}

def outpaint_image(image: Image.Image, prompt: str, directions: list, model_id: str) -> Image.Image:
    expand_top = 128 if "上" in directions else 0
    expand_bottom = 128 if "下" in directions else 0
    expand_left = 128 if "左" in directions else 0
    expand_right = 128 if "右" in directions else 0

    width, height = image.size
    new_width = width + expand_left + expand_right
    new_height = height + expand_top + expand_bottom

    canvas = Image.new("RGB", (new_width, new_height), (255, 255, 255))
    canvas.paste(image, (expand_left, expand_top))

    mask = Image.new("L", (new_width, new_height), 255)

    if model_id not in pipe_cache:
        device = "cuda" if torch.cuda.is_available() else "cpu"
        dtype = torch.float16 if device == "cuda" else torch.float32
        pipe_cache[model_id] = StableDiffusionInpaintPipeline.from_pretrained(
            model_id,
            torch_dtype=dtype
        ).to(device)

    pipe = pipe_cache[model_id]
    result = pipe(prompt=prompt, image=canvas, mask_image=mask).images[0]
    return result