For the upscalers: sharpness when the picture stands still (issue #8), and a test of crisper edges when it moves.
- "Sharpness at rest" (new slider under "Steady in fast motion", off by default). Reported: sharper while moving, softer when standing still.
Both are the same thing seen from two sides. Moving, the upscaled picture leans on a resample of the game's frame (sharper); at rest it is the
upscaler's own picture, and without the camera jitter a game gives its upscaler it has no extra detail to find, only its smoothing to give.
Measured on recordings with the frame held still, DLSS scores 32.5 dB against 33.6 for a plain resample, with 73 % of the game's detail against 80 %
(World of Warcraft: 58 % against 65 %). The slider keeps that share of the resample in the picture even where nothing moves (0.25: 32.9 dB,
0.5: 33.2, 1: 33.6), trading some of the upscaler's anti-aliasing at rest for sharpness; sharpening adds on top. Try 0.3 to 0.5. - "Crisp edges when moving (test)" (off by default): the resample the lean uses in motion can be AMD's FSR 1 edge-adaptive filter (EASU, ported to HLSL,
in NOTICE.md) instead of Catmull-Rom: cleaner, crisper thin and diagonal edges. Measured shimmer and detail are the same; the difference is in how
edges look, so it is there to compare by eye. On the host test's sliding picture it is closer for FSR (11.7 against 14.2 levels off). - Tools:
nr_srevaltakesrestmix=,still=1,lean=easu,sample=point(aliased input, as a game without anti-aliasing renders at a lower size),
stability=,fastp=,saveframe=, and prints the plain stretch beside its no-reference scores. Findings indocs/frame-generation-research.md:
the three upscalers side by side, how far the lean should go, and why a temporal upscaler cannot add detail without jitter.