vladmandic/sdnext

[Feature]: Add Pixelsmith support

Closed this issue · 4 comments

RedStingray commented

Feature description

Add Pixelsmith, an upscaling technique that uses cascading patches to reduce artifacts in the upscaled output.
Official code is here: Pixelsmith github
Link to paper is here: Arxiv page
Paper claims that it is state-of-the-art in both image quality and generation efficiency.

Version Platform Description

No response

vladmandic commented

author should make it portable instead of having hacks like this:

Copy/paste autoencoder_kl.py and vae.py in your environment (f.i. home/user/miniconda3/envs/pixelsmith/lib/python3.11/site-packages/diffusers/models/autoencoders/)

this is nonsense and i'm not going to do the work for them - if repo is updated so code is actually portable, i'm open to integrating it.

  • 👍2
RedStingray commented

Understood, thanks for looking at it!

renich commented

How about a mini-guide of how to use it somewhere?

Something like:

  • generate an image in txt2img at 1024 x 1024
  • go to img2img and use PixelSmith and use the generated image as base. Change the resolution to 2048 x 2048 and generate.
  • repeat but, now, with 4096 x 4096.

Is that correct?

vladmandic commented

closing the request as pixelsmith support has been added.

@renich yes, that is correct.

  • 👍1