Weighted Resize CRASH

[Selur]

Will look at it.
num_streams should be available independent from CUDA Graph and RT.
see: https://github.com/HolyWu/vs-femasr/blob..._init__.py

Cu Selur

[Dan64]

I noted that the parameter "trt" is missing in input in __init__.py.
It's strange because when I set it to true is generated in the folder a file like:
FeMaSR_SRX2_model_g.pth_NVIDIA GeForce RTX 3060_trt-8.5.2.2_608x320_fp16_workspace-33554432.pt

So RT should be supported...

P.S.
In effect it is available in the release 1.0 (not in the master)

[Selur]

TRT support was remove on 28.12.2022.
https://github.com/HolyWu/vs-femasr/comm...cc5243db1d
which might be related to the issues I reported, when using it, see: https://github.com/HolyWu/vs-femasr/issues/2
Didn't note the change myself, so I'll disable it in the next build, since it doesn't work properly.

Cu Selur

[Dan64]

Ok, I tried FeMaSR with no RT enabled.

It is able to produce much sharper images than realesr and with more details, but when the source is too bad also the result is quite bad.
You can find a comparison here:

https://imgsli.com/MTUxNzQ5

[Selur]

Happy that it works for your source, doesn't work on all sources that good.

Cu Selur

[andry62]

Dear Selur.
You write that you will disable TRT support. Please don't.
It works fine for me. There is really one drawback-
of the three TITAN RTX GPUs, it only works with one and it doesn't work
run three instances of Hybrid with different gpu id. But still
encoding speed is four times faster than with Fusion.
In general, the FeMaSR model is very interesting for sources
average quality. Defects in the form of flicker are easily removed.
So don't disable TRT support please.
Thank you.

[Selur]

Doesn't make sense to keep it, since it was already removed from the source.
https://github.com/HolyWu/vs-femasr/comm...cc5243db1d
Any update to the filter would cause crashes if the gui uses it,...
=> if you want to have tensor support you would have to convince HolyWu to add it back.

Cu Selur

[Dan64]

Hello Selur,

  I do agree with andry62. With 1 stream the RT support for FeMaSR is working quite well.
  I know there are some artifacts, but any AI resizer introduce artifacts and they can be mitigated with the weighted resizer, which you introduced with the last version.

  With RT disabled FeMaSR is too slow and is not practically usable.
  So please try to don't update FeMaSR, till the RT support will be added again in the master.

Thanks,
Dan

[Selur]

Not really keen on keeping an option which is known to cause issues.
I can keep the option available until a new release of FeMaSR is out, but I will update to a new version when it's available.
=> if you want to keep RT ask HolyWu whether he sees a way to fix it.

Cu Selur

[Dan64]

I read your post. You issue was that with FeMaSR there are "some strange highlights".
Moreover you specify that: "enabling and disabling nvfuser, cuda_graphs does not change these highlights".
So not seems that the problem is related to RT since both Fusion than cuda_graphs are disabled in Hybrid when RT is enabled.
Then HolyWu concluded that FeMaSR "is not suitable for tiling" then you concluded that "Tensor requires tiling to work" even if it seems that the problem that you reported is not strictly related to RT.
Probably there are problems even with RT, but the problem that you reported will not be solved by disabling RT.

Dan