Fixes
- Large videos no longer stall before generation starts. The check that decides whether a file can use the fast thumbnail path used to read the entire file — on a big movie that meant an extra full pass over it, and on large files over slower storage it could time out completely. When it timed out we assumed the worst and fully decoded the video, which is far slower. It now samples the file in a dozen places instead: same decision, a fraction of a second, and no dependence on file size. Measured on an 18 GB movie: 13 seconds down to 0.33 seconds, identical result.
- Files reported as "we couldn't read this file's snapshot frame layout" and then slow-decoded now generate at full speed. No need to remux them.
- Windows + NVIDIA is documented correctly — it works under Docker Desktop's WSL2 backend with
--gpus all. The docs previously claimed Docker on Windows couldn't use CUDA and pointed at an install-from-source path that never existed. - Working Windows installs are no longer told their GPU setup is "unofficial". That advisory fired even when GPU detection fully succeeded; it's now informational.
- In-app messages name the real Settings location — Settings → Processing Options → GPU Configuration, rather than the old "Settings → GPU".