Pixel histogram from GPU decoding shows unnatural trend and abnormal spikes

[entron]

Hi TorchCodec team,

When using VideoDecoder(..., device="cuda") to decode a large sample of frames from various videos, I observed the following issues in the grayscale intensity histograms:

1. The overall distribution is inconsistent with what’s expected from natural video content. The histogram lacks a smooth trend at the beginning and end.

2. Certain grayscale values show abnormally high spikes that do not occur when decoding using FFmpeg or OpenCV.

Here is the histogram from 10,000 frames decoded via TorchCodec on GPU:

If I set device to 'cpu' then the result looks more natural:

My guess for GPU issues:

1. TV ranges are mapped to PC range for some video but not for the others (e.g. when color_range=unknown)
2. Probably related to 10 bit video decoding and how it is converted to 8 bit (see also: noise when decoding hevc yuv420p10le videos with device="cuda" #598 ). I do not see the high spikes for 8-bit videos.

Thanks for looking into this.

[NicolasHug]

Hi @entron , thank you for the report and for providing these insightful histograms. We will be looking into this.

[NicolasHug]

Also can you share the FFmpeg version you were using? We've observed pretty bad results on FFmpeg4 with CUDA decoding, which I think are due to the underlying codec implementation

[entron]

@NicolasHug This is my version:

ffmpeg version 7.1.1 Copyright (c) 2000-2025 the FFmpeg developers
built with gcc 13.3.0 (conda-forge gcc 13.3.0-2)

[NicolasHug]

Quick update: from #777, we are now raising a loud error on such 10bit videos that we cannot decode properly. I think we'll soon be able to support those though, so stay tuned!

Thank you for the report, I'll close this issue as we are now avoiding silently wrong results. We're tracking 10bit videos support efforts in #776