HMA (Current #1 Upscaling Leader)

[Arcitec]

The models were finally released 3 weeks ago! Dear God, please... support this! :D

Leaderboard Ranking:

• https://paperswithcode.com/paper/hmanet-hybrid-multi-axis-aggregation-network

Repo:

• https://github.com/korouuuuu/hma

Models:

• https://drive.google.com/drive/folders/13Bxt_BXtWgwwcO7tjfoL6zYZAUY1v3zi?usp=sharing

[Arcitec]

Important: The HMA code is based on HAT's code. It was a direct copy of HAT's repo, with some code rewrites and better training.

You already support HAT, so adding HMA support will be a lot easier than I expected.

However, there's problems in HMA's code. I have documented all the steps I had to do to fix the bugs in HMA's / HAT's repos to make the code run locally. I submitted a PR, so it's fixed now. Just follow these steps if you want to run it locally directly from their repo:

korouuuuu/HMA#4 (comment)

I think the best path forward to figure out how to implement HMA is to do a diff between HAT and HMA's repos (possibly at an older point in time of HAT's repo since I doubt HMA used the latest HAT code), and see what part of HAT's architecture that HMA actually changed. You probably have 99% of the code already.

[Arcitec]

@the-database Hi, since I saw that you have a huge interest in upscalers and have contributed a lot of pull requests for various upscalers, I am curious if you were aware of HMA?

It's still ranked #1 in tons of image categories (search for #1 here: https://paperswithcode.com/paper/hmanet-hybrid-multi-axis-aggregation-network).

The results are mindblowing.