AC3 to AC3 → ffmpeg crash

[ToiletDuck]

You could trigger: "Config->Tools->Tool Versions" on both releases and compare the output.

---------------------------
Info
---------------------------
Known tool versions:
  16bit-x265: 1.6+360-bca33880585a
  aomenc: 3.8.0-131-g77657197c7 (default)
  avs2yuv: 0.24bm6
  avs2yuv64: 0.28
  AVSMeter: 3.0.9.0
  AVSMeter64: 3.0.9.0
  bdsup2sub++: 1.03
  dcaenc: Unknown
  delaycut: 1.439
  DivX265: 1.5.0.8
  faac: 1.30
  fdkaac: 1.0.5
  ffmpeg: N-112974-gd324415563-g5256b2fbe6+2
  ffmpeg_32: N-112879-g444806f211-gd0ec826077+1
  FFVideotoolbox: no version info
  fhgaacenc: 20120806
  flac: c - Command-line FLAC encoder/decoder version git-72787c3f 20231124
  flaccl: 2.2.3
  FLVExtractCL: 1.6.2
  kvazaar: 2.2.0-15-g0d6651bf 2023-12-13
  lame: 3.100
  lsdvd: no version info
  MediaInfo: 23.11
  mencoder: 38359
  mencoder_32: 38359
  mkvextract: 81.0
  mkvinfo: 81.0
  mkvmerge: 81.0
  MP4Box: 2.3-DEV-rev696-gf8a41b2ae-ab-suite
  mp4fpsmod: 0.26
  mplayer: 38359
  mplayer_32: 38359
  mpv: no version info
  mpv_32: no version info
  neroAacEnc: 1.5.4.0
  NVEnc: 7.4
  NVEnc (FFmpeg): no version info
  oggenc: 1.4.2
  opusenc: 0.2-34-g98f3ddc
  qaac: 2.79
  QSVEnc: 7.57
  rav1e: AV1 video encoder
  sox: 14.4.2
  SVT-HEVC: 1.5.1
  telxcc: 2.6.0
  tsMuxeR: 3.1
  VCEEnc: 1
  vpxenc: 1.13.1-599-gdf655cf4f
  vspipe: R65
  vStripFree: 0.8e
  wine: no version info
  x264: 3172
  x265: 3.5+113-3fa1946b3
---------------------------
OK  
---------------------------

I know you can do it like this ↑..  Iam kidding back there..

It's just that i have to remember and not forgot to take notes like that, before installing each & new release.. 
Hence, it's easier for me to remember wich tool version has changed,  if i could read it from a log that's available with each new release.

Talking of wich, do you know if the Resize filter "ESRgan" gets frequently updated ?
i have a feeling , i'll make good use of that filter for my dvd collection in the future.. xD

That filter is a new & pleasant experience for me..  It realy is an good alternative for denoisers like KLNmeans, wich produces very same results.  But ESRgan, realy does a outstanding job at keeping details / sharpness crispness of the image , wich knlmeans lacks imo.. 

I know nlmeans is more of a denoiser, while esrgan is essentially a resizer / upscaler.  But i have to say, that esrgan also have a denoise setting, wich does a good job at removing noise while keeping details though .
That combined with an good color filter for gama (black / white) adjustings.. Looks awesome, almost HDR i'd say wow.. 


Only i REALY need an POWERHOUSE GPU .. 
iF ONLY i could utilize my cpu more... cpu utils only like 10 - 15% , and oddly enough the gpu utilizes only for about 55 - 65% when i use ESRgan ? 

Is that about right?
Maybe that will change if i'll use studio drivers ..

anyway, tHanks for the reminder..

cheers,
td

[Selur]

Talking of wich, do you know if the Resize filter "ESRgan" gets frequently updated ?

No clue what you are referring to.
ESRGAN stands for Enhanced Super-Resolution Generative Adversarial Networks, which is a pipeline using (training) models for super resolution tasks.
Hybrid allows to use:

• RealESRGAN (https://github.com/HolyWu/vs-realesrgan)
  
• VSGAN which supports ESRGAN models (https://github.com/rlaphoenix/VSGAN) of which Hybrid contains a few in the vsgan_models folders, other models can be found over at https://openmodeldb.info/ (not all models there are supported by VSGAN)
  
• VSMLRT which supports ESRGAN models (https://github.com/AmusementClub/vs-mlrt) in onnx-format of which Hybrid contains a few in the 'vs-mlrt/models'-folders, other models can be found over at https://openmodeldb.info/ (not all models there are supported by VSGAN, but some can be converted using chaiNNer).
  

the first two get updates with the torch-addon, the last one gets updated with the mlrt-addon, the models get updated by the corresponding packages.
I keep not track of when I update which.

iF ONLY i could utilize my cpu more... cpu utils only like 10 - 15% , and oddly enough the gpu utilizes only for about 55 - 65% when i use ESRgan ?

gpu utilization depends on:
a. the gpu and drivers
b. the model used
d. the configuration of the filter, like the number of parallel streams used whether tiling/fp16/fp32/Fusion/TensorRT/.. are used.
If you do not understand what an option does, read up on it.

Cu Selur

[ToiletDuck]

Talking of wich, do you know if the Resize filter "ESRgan" gets frequently updated ?

No clue what you are referring to.
ESRGAN stands for Enhanced Super-Resolution Generative Adversarial Networks, which is a pipeline using (training) models for super resolution tasks

Iam talking exactly about that ↑ → Models/data esrgan uses ... 
If iam right the last update dates from march 2023 so it seems   .  

Presumable, the maximum potential of the filter has been reached.

iF ONLY i could utilize my cpu more... cpu utils only like 10 - 15% , and oddly enough the gpu utilizes only for about 55 - 65% when i use ESRgan ? 

gpu utilization depends on:
a. the gpu and drivers
b. the model used
d. the configuration of the filter, like the number of parallel streams used whether tiling/fp16/fp32/Fusion/TensorRT/.. are used.
If you do not understand what an option does, read up on it.
Cu Selur

about:

D. i think i have maxed out everything in favor of performance when using ESRgan and my current GPU.  As suggested, iam using TRT + FP16, 1 stream (more = gpu taxing !) ..
What about tiles x/y ?  would that improve speed, and at what setting is there noticeable quality loss in your experience ?  And then there's the setting WSS = xxxxxxxxx  !

I guess i can use a test sample to figure it out, but i like to know whats your experience is using different settings.


EDIT: About stream setting in Esrgan i get the following script error ↓

Failed to evaluate the script:
Python exception: realesrgan: setting num_streams greater than `core.num_threads` is useless

Traceback (most recent call last):
File "src\cython\vapoursynth.pyx", line 3115, in vapoursynth._vpy_evaluate
File "src\cython\vapoursynth.pyx", line 3116, in vapoursynth._vpy_evaluate
File "D:\TEMP FOLDER\tempPreviewVapoursynthFile16_14_20_845.vpy", line 60, in
clip = RealESRGAN(clip=clip, model=5, device_index=0, trt=True, trt_cache_path=r"D:\TEMP FOLDER", denoise_strength=0.25, num_streams=2) # 2880x2304

cheers
TD

[Selur]

am talking exactly about that ↑ → Models/data esrgan uses ...

↑ → means nothing to me.
Since you do not use the names used in Hybrid, I have no clue what you are talking about.

1 stream (more = gpu taxing) ..

Whether more streams speed up or slow down stuff depends on the models, filters, the vram requirements per stream etc.

What about tiles x/y ? would that improve speed, and at what setting is there noticeable quality loss in your experience ?

tiles rarely increases speed, it's mainly meant to lower vram usage at the cost of speed
whether tiling results in quality loss depends on the tiling size and the used model.

Cu Selur

[ToiletDuck]

am talking exactly about that ↑ → Models/data esrgan uses ...

↑ → means nothing to me.
Since you do not use the names used in Hybrid, I have no clue what you are talking about.

I have  the feeling we are running in ↑→↓← ciricles here..

If you don't know what i mean, i don't know it either anymore i guess..

It's my understanding (correct me if iam wrong) that the filter make use of certain →pre-trained models ← / data as a reference point.  
And i did read that exactly that ↑ yes ↑ exactly that wos updated in the past by the developer.  For the users to add/replace it manually ..

iAM talking in general, not specific what model..

But if iam still not clear to you, nevermind my faq..

cheers,

[Selur]

Python exception: realesrgan: setting num_streams greater than `core.num_threads` is useless

Makes sense.
With "num_streams=2" this should only happen if you limit the number of cpu threads to 1.

Cu Selur

[ToiletDuck]

Python exception: realesrgan: setting num_streams greater than `core.num_threads` is useless

Makes sense.
With "num_streams=2" this should only happen if you limit the number of cpu threads to 1.

Cu Selur

that's not what i meant..

But the "stream" setting in ESRgan is pretty much useless indeed, and offers no benefits wotsoever imo..  
It's quite the opposite.. 

[Selur]

Have no problem using it here and depending on the model it speeds things up. (at the cost of vram)

[ToiletDuck]

Have no problem using it here and depending on the model it speeds things up. (at the cost of vram)

You sure about that ↑ ?

Depending on the model ?  so, setting the stream to only 2 or in your case 4 using the model Realesr-general for example improves the speed for you ?

When i set it to only 2, my system doesn't move at all anymore... i'm inclined to think it even goes backwards >_>

And it's not because of the vram either, it uses like only 25% vram in my case ..!


EDIT: 

I almost forgot to ask, but you've mentioned something about limiting the cpu threads to 1..  Are you talking about a setting in hybrid that does that ? If so, where that at ?
Iam not talking about taskman or any other processmon tool to do that .. fyi.

[Selur]

Depending on the model ? so, setting the stream to only 2 or in your case 4 using the model Realesr-general for example improves the speed for you ?

Did a quick test with 'realsr-general' and a sd source: no
But, Yes, it did help with other models. (using VSGAN, VSMLRT)
I rarely use RealESRGAN or other machine learning upscalers, since doing ml based denoising and restoration and then upscaling with nnedi3 usually gives me better results.
=> test it for yourself. Dependent on the source, the model used the other settings you might get different results.
Might be that it doesn't help with RealESRGAN at all, or even hurt.

i'm inclined to think it even goes backwards

That could happen. Since it only helps if the gained speed over-compensates the additional needed syncrhonization.

Cu Selur