search

rlaphoenix / VSGAN

PyTorch-based Super-Resolution and Restoration Image Processing Module for VapourSynth
https://vsgan.phoeniix.dev
MIT License
190 stars 13 forks source link

New pip release? #43

Open Selur opened 8 months ago

Selur commented 8 months ago

Seeing that 1.6.4 from Jan 25, 2022 is still the current version served by pip and there have been quite a few changes since then, I was wondering if there are any updates for a new pip release are planned. :)

evolutionaryman commented 7 months ago

I also am curious if this will get updated. :)

rlaphoenix commented 5 months ago

Hi, a lot of the commits made after the last release I'm currently unsure-of as in if they actually make a difference good or bad, hence I haven't yet made a stable release for. If you guys don't mind testing it out and seeing if you find any problems, it may speed up release. The main issues I'm concerned with atm is if there's any loss to support for any input/output bit depth or format combinations, and any speed or memory usage changes.

Selur commented 5 months ago

I can do some testing next week. Anything a bit more specific to test? I assumed, so far, input needed to be: a. mod4 b. RGBS or RGBH c. pc luma scale

rlaphoenix commented 5 months ago

With VSGAN, afair (I havent worked on it or used it in ages), it supports any combination of input format, so long as it wasnt YUV.

Selur commented 5 months ago

Okay, I tested compatibility only atm.: input: RGB24, RGB27, RGB30, RGB36, RGB42, RGB48, RGBH, RGBS 2-3 different models per architecture with different scalings. (for EGVSR. I only used: EGVSR_iter420000.pth) tested with: vsgan = ESRGAN(clip=clip,device="cuda") => no problems vsgan = SwinIR(clip=clip,device="cuda") => no problem vsgan = HAT(clip=clip,device="cuda") => no problem vsgan = EGVSR(clip=clip,device="cuda") => broken, different colored and wrong output for the tested RGBXX inputs.

Selur commented 5 months ago

Ran a few speed comparisons with different ESRGAN models and the new one seems to be slightly faster (~1%) for normal models, but a bit faster (~7%) for compact models. I didn't test EGVSR and the color changing depending on the input is also broken in the old version, but the output wasn't as broken. => EGSVR doesn't seem to work properly.