Selur's Little Message Board

Not sure if you had added these models to the dev version. It looks promising. If it isn't implimented I would love to test these out with the hybrid pipeline.  


https://github.com/xpixelgroup/hat


Here is where I learned about it:

https://paperswithcode.com/task/image-su...testPlease, read the 'Infos needed to fix&reproduce bugs,..'-sticky before you post about a problem.

this:


is stunning!

Not sure what models were used to obtain these results.
I tried to replicate the test using the resizers available in Hybrid: BasicVSR++, RealESRGAN (x4plus), RealESRGAN (realesr-general)



Then I added improved version by using: CASglis, FastLineDark, CAS (after resize) and weighted resize with weight = 0.65

I was surprised by the behavior of  RealESRGAN (x4plus) on the dog image, because the resized image was too smooth, this effect was mitigated using the  weighted resize with weight = 0.65.
In general RealESRGAN with realesr-general seems to provide better results.

Did you test these through VSGAN or vs-mlrt?

Did a quick test.
Those models require additional parameters, so your whishes needs to go to https://github.com/rlaphoenix/VSGAN and/or https://github.com/AmusementClub/vs-mlrt/ to support these models.

Cu Selur

(17.12.2023, 19:37)Selur Wrote: [ -> ]Did you test these through VSGAN or vs-mlrt?

None of them, I used "Frame->Resizer->RealESRGAN" and "Frame->Resizer->BasicVSR++"
My startup log is the following



Dan

---

(17.12.2023, 20:28)Selur Wrote: [ -> ]Did a quick test.
Those models require additional parameters, so your whishes needs to go to https://github.com/rlaphoenix/VSGAN and/or https://github.com/AmusementClub/vs-mlrt/ to support these models.

Cu Selur

I noted that VSGAN added the support to HAT, I found the models in GoogleDrive, which parameters have to be passed ?

If those models can be used through VSGAN or VSMLRT (both listed as resizers, first uses .pth the second .onnx models) adding them is easy, by simply loading the model.
If those models can not be used through these, either they need a whole wrapper filter (like BasicVSR++) or would need adjustments to VSGAN or VSMLRT.

Cu Selur

Ps.: just checked, the latest VSGAN version available through pip is still 1.6.4 which is from ~2years ago, so unless a new version is released, I doubt that those models will work.

(18.12.2023, 19:52)Selur Wrote: [ -> ]Ps.: just checked, the latest VSGAN version available through pip is still 1.6.4 which is from ~2years ago, so unless a new version is released, I doubt that those models will work.

The support to HAT was added 7 month ago. Any way to apply the latest github version ?

Don't know for sure, I never tried, but in theory, updating the files in
and removing the existing __pycache__-folders should work.

Cu Selur

Ps.: I created an issue entry https://github.com/rlaphoenix/VSGAN/issues/43

(18.12.2023, 20:05)Selur Wrote: [ -> ]Ps.: I created an issue entry https://github.com/rlaphoenix/VSGAN/issues/43

Good question