upscale_image.py

import torch
from diffusers import AutoencoderKL
from diffusers import StableDiffusionUpscalePipeline
from PIL import Image

device = "cuda"
seed = 100
def execute(input_image):
    model_id = "stabilityai/stable-diffusion-x4-upscaler"
      # GPUを使用する場合
    pipe = StableDiffusionUpscalePipeline.from_pretrained(model_id, torch_dtype=torch.float16, variant="fp16") #not support vae?
    pipe = pipe.to(device)
    pipe.enable_attention_slicing()
    #pipe.enable_model_cpu_offload()
    pipe.enable_xformers_memory_efficient_attention()
    #pipe.enable_vae_tiling()
    pipe.vae.enable_tiling()
    # 画像のパスとプロンプト
    prompt = "beautiful girl"
    first_resize_w = 0
    first_resize_h = 0
    
        # 画像の読み込みとリサイズ
    image = input_image#.convert("RGB")
    low_res_img = image
    if first_resize_w!=0 and first_resize_h!=0:
        low_res_img = image.resize((first_resize_w, first_resize_h))
        
    upscaled_image = upscale(pipe, prompt, low_res_img)
    return upscaled_image

def upscale(pipe, prompt, img, step=50, guidance_scale=7.5):
  generator = torch.Generator(device).manual_seed(seed)
  return pipe(prompt=prompt,generator=generator, image=img, num_inference_steps=step, guidance_scale=guidance_scale).images[0]


if __name__ == "__main__":
    image = Image.open("sample.jpg")
    upscaled_image = execute(image)
    upscaled_image.save("output.jpg")