Performance - prnvkhndlwl/Kinetrace GitHub Wiki

Because decoding is the bottleneck, a faster GPU does little; these help:

Frame cache — sized automatically to a quarter of your RAM (max 24 GB), so a whole working region of 4K frames stays in memory and re-scrubbing a stretch you are correcting is instant. Override with the KINETRACE_CACHE_GB environment variable if you want more or less.
Decode backend — hardware decode roughly doubles sequential decode (long tracking runs) but is not automatically better for seeking. Measure your own footage, then apply the winner:
```
.venv\Scripts\python.exe tools\bench_decode.py "D:\path\to\your4k.mp4"
```
Apply with KINETRACE_DECODE = auto (default) | msmf | hw | ffmpeg. Pick one and keep it for a project: backends agree on frame numbering but differ marginally in color conversion.