CutBench reads a long gameplay recording or a link to your own past Twitch or YouTube broadcast and returns the best moments as vertical 9:16 clips, ranked best first, with captions burned in.
It runs on your own computer. Footage is read from disk and written back to disk. Nothing is uploaded for processing. Captions follow the language spoken in the recording, Japanese included, and are not translated. A .srt file is saved next to each clip.
Optional publishing: scheduled upload to YouTube Shorts, and TikTok delivery to your drafts.
Requirements: Windows 10 or 11, NVIDIA GPU with 6 GB VRAM and CUDA, 8 GB RAM, 30 GB disk. AMD and Intel Arc are outside the supported spec: the work runs on the processor, tens of times slower, and still completes. Measured on one RTX 3080 over 15 runs, a job takes 0.6 to 1.1 times the length of the recording, plus 5 to 10 minutes. It is a batch tool, not a live clipper or a screen recorder.
One purchase on Steam, no subscription. Exports carry no watermark.
Not to be confused with the CutBench benchmark from the Cut2Next research paper.
No comments or reviews, maybe you want to be first?