State of the project: v3.31.0

[SysAdminDoc]

Quick orientation for anyone landing here fresh. This began as a fork of YaoFANGUK's video-subtitle-remover and has since been rebuilt around a different goal: results you can trust unattended. The pieces that matter are real inpainting (LaMa through ONNX Runtime, no torch install required), temporal background exposure that reconstructs the actual background from neighbouring frames instead of hallucinating one, and scene-aware routing that picks between STTN and ProPainter per segment.

3.31.0 finished something I'd wanted for a long time: every queue job now runs supervised in its own child process. A crash mid-batch takes down one job, not your overnight queue. Alongside that, WebVTT now translates as WebVTT with an explicit report of anything lossy, and you can freeze an approved matte and reuse it as a fail-closed input for later jobs on the same source.

Where help is genuinely useful: hardware coverage. The backend matrix (CUDA, DirectML, CPU fallback, plus the hardware decode hints) has more combinations than one person can test. The thread pinned next to this one asks what you're running and what still fails. Two minutes of your time there directly shapes what gets fixed first.

Questions go here, bugs go in Issues with the template. If you've processed something interesting, Show and tell exists and I do read it.

[spartian137-jpg]

Hi, I am probably the dumbest person here but is there a guide video where it runs you though putting in a video and having the software translate and impaint the subtitles?