How to do the equivalent in this repo?

[alexblattner]

Hello @victorca25, I'd like to replicate highres fix, but am unsure how it works exactly. I mainly use diffusers so I must use your repo to replicate highres fix.

If I wanted to use 4x-UltraSharp, upscale it by 1.5, with 25 highres steps and a denoising of 0.4. What should the code look like?

this is how the default code looks like:

import torch
from PIL import Image
import numpy as np
from run import load_model, run

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

model = load_model('upscaler/4x-UltraSharp.pth', device)

path_to_image = 'img11.png'
image = Image.open(path_to_image).convert('RGB')
 
sr_image =run(model,image)
pil_image = Image.fromarray(sr_image)
pil_image.save('u3.png')

I hope you see my issue soon and thanks in advance for your time.

[victorca25]

Hello! Did you find a solution to your question?

The original idea with this repo was to run a model with many images in a directory directly from the command line, but making a script like the one you want to use for the Hires Fix with diffusers would be very simple, since it happens on the pixel-space.

[alexblattner]

@victorca25 I haven't. I just upscale by 4 then shrink it as I need. I don't know enough about "since it happens on the pixel-space."