The upscalers steady in fast camera turns.
- Steady in fast motion (all three upscalers). Lossless Scaling's frames carry none of the sub-pixel camera jitter a game gives its
upscaler, so in a fast turn the upscaler's history adds little and trails: leaves and edges with soft doubled outlines. Now, where the
picture moves fast (from 0.5 % of its width a frame) or its motion cannot be followed, a pass after the upscaler blends its picture toward
this frame upscaled plainly. DLSS needed it most: its DLSS 4 models ignore the mask a game would give it for this. Measured offline on a
recorded Silent Hill f turn, shrunk 1.5x and upscaled back (at a quarter of the size, where trailing shows; a plain stretch 43.3 dB):
DLSS 36.8 -> 44.2, FSR 41.7 -> 45.6, XeSS 40.2 -> 46.2; all three now beat the plain stretch in the turn, and a slow pan is unchanged.
On by default; "Steady in fast motion" in each upscaler's panel turns it off to compare. XeSS also gets the motion's distrust mask now
(as its responsive mask), as FSR and DLSS always did. - The DLSS model list, one line each: K (NVIDIA's default), L (recommended), J, M (DLSS 4.5) and E (DLSS 3's CNN). On the same
recording L is the only one that does not drift from the picture in a slow pan, and close to the best in fast turns; E is the best in
fast turns and the lightest; M the weakest here and the heaviest. - The motion estimate on a character in a fast turn. A character the camera follows (dark, faintly textured) was held to the guess of
the background sweeping past and took random vectors; the pull toward the guess is now capped, so it finds its own. For the upscalers,
Neural Rendering and frame generation alike. - Neural Rendering's Temporal smoothing no longer trails in a fast turn: the last frame's change is held within the range of this
frame's own around each pixel (as temporal anti-aliasing does), and the smoothing fades out in fast motion. (Off by default.) - HDR: the frame's light is clamped at 0 before the upscalers (an HDR10 colour outside Rec.709 came out negative).
- DLSS Upscaler and Neural Rendering together: switching Neural Rendering off while the DLSS Upscaler ran shut NVIDIA's NGX down under
it ("FeatureNotFound" in the log) and the upscaler switched itself off. NGX now shuts down only when the last of our addons using it
does, and the DLSS Upscaler starts NGX again if its feature is lost anyway. - Frame generation of our own (the FSR Upscaler; a prototype, not in this release: only in a test build, CMake NR_FRAMEGEN=ON) from its first live tests: FSR 3.1 frame generation is
dispatched directly (its own swap chain paced itself and, with our pacing, halved the real frame rate, 30 -> 18); the frame between and
the real one go out evenly spaced; HDR frames are measured in their SDR view; a guard takes back what FSR pastes over a character in a
turn ("Keep the character clean in turns"); what it did is logged every 10 s. - The recorder can keep what is shown ("Record what is shown", with frame generation of our own, in a test build): the frames at the screen's size,
the frames made between and the real ones, each marked (nr_lsrec exportnames them). - Tools:
nr_srevalscores an upscaler offline on a recording (the addon's own engine: DLSS, FSR or XeSS, any setting);nr_fgeval
gains the live gap, a score around the character, FSR's debug views, direct dispatch and the addon's own frame generation engine.
Findings indocs/frame-generation-research.md. - Tests: targeted runs of the upscaler and Neural Rendering scenarios (moving pictures on all three upscalers, sharpening, edges, HDR,
smoothing), and offline measurements withnr_sreval(now with a steadiness score: shimmer and flicker from frame to frame) and
nr_fgeval. Live: the DLSS Upscaler in World of Warcraft (the owner: happy with it). Not yet tried live: Neural Rendering's changes and
the NGX fix above. - Known: some flickering while moving with the DLSS Upscaler in World of Warcraft (noted, to look into).