こすたろーんエンジニアの試行錯誤部屋

作成物の備忘録を書いていきますー

【Stable Diffusion】try to make a high resolution image with Diffusers upscaler

スポンサーリンク

I have used diffusers and LoRA to generate various images.
technoxs-stacker.hatenablog.com

This time I tried to see if I could create a high resolution image
This article is just a reminder, but I hope it will be useful for everyone!

contents

スポンサーリンク

abstract

How to use diffusers and upscaler on Google colab to create high resolution images.

1.requirement

Google Colab
Diffusers transformers==4.26.0
model : chilloutmix_NiPrunedFp32Fix
upscaler:stabilityai/stable-diffusion-x4-upscaler

2.code

from diffusers import StableDiffusionUpscalePipeline

def upscale(pipe, prompt, img, step=100):
  return pipe(prompt=prompt, image=img, num_inference_steps=step, guidance_scale=0).images[0]

# make pipeline  
pipe_upscaler = StableDiffusionUpscalePipeline.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", torch_dtype=torch.float16, revision="fp16")
pipe_upscaler = pipe_upscaler.to("cuda")

prompt = "prompt text"
im = Image.open(save_path).convert("RGB")#save_path:path to image
low_res_img = im.resize((256, 256))
step = 200
pipe_upscaler(prompt=prompt, image=low_res_img, num_inference_steps=step, guidance_scale=0).images[0]

3.result

normal(552 x 552 pixel)

high resolution(1024 x 1024 pixel)

The image is clearer after the high resolution when compared to the enlarged image.
left:normal / right:high resolution

スポンサーリンク

refarence

huggingface.co github.com note.com
huggingface.co