Selur's Little Message Board

It's weird, and I'm not sure it's bug or something wrong about my system. Anyway I'll describe what happened at below:

I always use Hybrid in portable configuration.

Last night I tried to update&save new global profile with a new setting I never saved before: VS yuv420 chroma location left.

After update the profile, I restart hybrid, and suddenly, DPIRdenoise(mlrt) was always skipped(Hybrid runs like it was not enable), SCUNet(mlrt) always crashed straightaway(generate a 5kb output file). and now I can't run vs scripts which first running in RGB plane, then later convert to yuv plane with "yuv420 chroma location left" setting on. The scripts always show:

Quote:# adjusting color space from YUV420P8 to RGB24 for XXXXXXXXX
clip = core.resize.Bicubic(clip=clip, format=vs.RGB24, matrix_in_s="709", range_in_s="limited", chromaloc=0/)

The "/" at the end looks weird, and the job crashed immediately just like SCUNet(mlrt).

The weirdest thing is that the portable hybrid suddenly not being portable anymore:

I delete the whole hybrid dir and reinstall hybrid with latest dev and make it portable again, then reinstall the latest windows addon, but situation remains. Also tried running in installed configuration / uninstall it & delete the whole settings,  then running another new portable config, both still the same.

Windows Hybrid_dev_2026.06.19-214655
Avisynth32bit_2024.12.15
onnx_models_2023.12.05
VapoursynthR77_torch_2026.06.16
vsgan_models_2023.07.10
vs-mlrt_2026.06.06

2 questions BTW

1. Does hybrid dev always produce a minidump at C:\Users\USERNAME\AppData\Local\CrashDumps even if nothing happened(only launched it and closed immidiately)? (no minidump with Hybrid_2026.03.21-1)

2. Is there any matching issue with Hybrid version & addon version? For example, the location of vspipe.exe in VapoursynthR73_torch_2026.02.26 / VapoursynthR77_torch_2026.06.16 is different, can I change between them with the same Hybrid version?

Yeah, that's a typo, I made during refactoring.
=> uploaded a new dev which should fix the issue.

Out of curiosity: Why are you still installing Avisynth 32bit?

Quote:1. Does hybrid dev always produce a minidump at C:\Users\USERNAME\AppData\Local\CrashDumps even if nothing happened(only launched it and closed immidiately)? (no minidump with Hybrid_2026.03.21-1)

Seems like it.
I'll investigate; this is not intended. (will report back later)

Quote:2. Is there any matching issue with Hybrid version & addon version? For example, the location of vspipe.exe in VapoursynthR73_torch_2026.02.26 / VapoursynthR77_torch_2026.06.16 is different, can I change between them with the same Hybrid version?

older Hybrid versions will not work properly with newer R77 torch-addons.
If you remove the old Vapoursynth folder and replace it with the add-on newer can use older add-ons, but might crash if a filter call changed.

Cu Selur

Think, I fixed the CrashDump-bug and uploaded a new dev (Hybrid_dev_2026.06.20-20172).
Let me know whether that version fixes the issue for you too.

Cu Selur

Quote:Out of curiosity: Why are you still installing Avisynth 32bit?

I didn't know that I didn't have to 


Tested Hybrid_dev_2026.06.20-20172:

chromaloc=0 issue fixed

minidump issue not fixed
DPIRdenoise(mlrt) & SCUNet(mlrt) issue not fixed

Noticed a new behaviour: If only enable DPIRdenoise(mlrt), Hybrid will run like it was skipped (90 fps encoding speed), but if I add another filter such as CAS behind DPIRdenoise(mlrt), now the job will finished immidiately like SCUNet(mlrt), with only 5kb output file.

And another bug I found:

If I load a input file, and load it again after finish loading, a window box will show:

Quote:Error
---------------------------
Script output (25) and expected fps (23.976) differ!

Quote: minidump issue not fixed

Are you sure, you used Hybrid_dev_2026.06.20-20172:
I can't reproduce it with this version anymore.

Quote: DPIRdenoise(mlrt) & SCUNet(mlrt) issue not fixed

Side note: if you are using "TensorRT RTX" Backend: that is known for be buggy in vs-mlrt, try TensorRT or DML (GPU) instead.
Also, 90 fps doesn't have to be wrong,.. (I get ~130fps for DPIRdenoise(mlrt) and ~50fps for SCUNet(mlrt) on SD content)
1. Check the Vapoursynth Preview with different Backends
2. Create and share a debug output for each of these issues.

Quote:, with only 5kb output file.

That should cause Hybrid to crash the job.

Quote: If I load a input file, and load it again after finish loading, a window box will show:

Can't reproduce that either.
-> create a separate debug output for that too. I probably will also need a proper step-by-step of what you are doing and a small sample video to test.

Cu Selur

Ps.: going to bed know, but I can look at it tomorrow

Extra questions:

1. Assuming there are 2 different version of portable Hybrid on disk, compatible to the same addon versions, can I move the addon files between them, or is it possible to share 1 set of addon files between 2 portable Hybrid?

2.Can I use the same engine file path between 2 different version of portable Hybrid with same version of addons? Or  can I use the same engine file path between 2 different version of portable Hybrid with different version of addon?

Regarding 1:
Sharing is a bad idea, since they would then share the same Python environment, which would cause issues.
Copying should work (delete __pycache__ folders), assuming both versions support these addon and you delete the Vapoursynth folder before copying.

Regarding 2:
Hybrid itself does not care about the engine files. Only vs-mlrt and torch do.
Different add-on versions will need different engine files. So if you have different version, you should use different engine folders.
I suspect that if you use the same add-on version, you can use the same engine files. (Not sure, test it, if you want to know for sure. I suspect it might work as long as not both versions use the same engine file,..)

About your DPIR and ScuNet crashes: try deleting existing engine files.
vs-mlrt does not always properly detect older/newer engine files.

In general I would recommend to keep everything separated between multiple portable versions.

Cu Selur

file loading issue debug output[attachment=3654]

minidump issue debug output[attachment=3661][attachment=3662]

Encoding with DPIRDenoise(mlrt)(Tensor RT) enabled only, then disable it and encode again. Generating same size output file, no engine file generated, and both running at same speed up to 200 fps for 1920x1080 30 sec bluray source input, which means dpirdenoise was not working at all [attachment=3664]

Encoding with DPIRDenoise(mlrt)(Tensor RT) + CAS. now it works (so strange), but at very slow speed (3~4 fps), it was 15 fps before issue happened. [attachment=3665]

Encoding with DPIRDenoise(mlrt)(Tensor RT "FP16") + CAS. Not working again, job finished immediately with 5kb output file, and no engine file generated. [attachment=3666]

Encoding with SCUNet(mlrt)(TensorRT). Works ok at 1.5 fps. Was up to 3~4 fps before. [attachment=3667]

HybridDebugOutput-LoadingFileErrorWindowBox(DebugEnabledByDefault) -> I managed to reproduce the problem.
But not sure atm. why this happens.
=> looking into it