Use nlmeans and FSRCNNX 16 for the Quality upscaling tier

ArtCNN is tuned for animation and looked unnatural on live action, and
its first Direct3D compile took about a minute per model. Quality now
denoises with AN3223's nlmeans (general-purpose profile, any size) and
upscales with FSRCNNX 16-0-4-1, trained on general imagery, at any
enlargement. Both compile in seconds; 1080p in a maximized 1440p window
measured 3.2 ms per frame.

The first-compile notice now covers any shader set not yet compiled on
the machine but only shows after three seconds, and a set with nothing
to verify at the current size clears it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Elijah 2026-10-30 19:51:47 -07:00
parent 506c69b36e
commit 5b6c063ff1
9 changed files with 3941 additions and 7897 deletions

View file

@ -150,5 +150,5 @@ export const UPSCALING: Record<Upscaling, { label: string; detail: string }> = {
off: { label: 'Off', detail: 'mpv’s standard scaling' },
performance: { label: 'Performance', detail: 'Sharp scaling, debanding and light sharpening' },
balanced: { label: 'Balanced', detail: 'Neural upscaling (FSRCNNX), colour reconstruction, debanding and sharpening' },
quality: { label: 'Quality', detail: 'Larger neural upscaler (ArtCNN) with denoising, colour reconstruction, debanding and sharpening. Needs a capable GPU' },
quality: { label: 'Quality', detail: 'nlmeans denoising, larger neural upscaler (FSRCNNX 16), colour reconstruction, debanding and sharpening. Needs a capable GPU' },
};