The upscalers in motion: a very big improvement. Less trailing and smearing while you walk and turn, no flicker at 1:1, and all three
upscalers now beat a plain stretch in a fast turn, for 3-4 ms a frame at 4K.
-
The upscalers lean on the new frame from much less motion. Lossless Scaling gives an upscaler none of the sub-pixel camera jitter
a game gives it, so in motion its memory of earlier frames adds no detail, only trailing. Measured offline, it trusted our motion
vectors and kept that memory: with them it did worse than with none. So the upscaled picture now leans on the new frame from about
1 px of motion a frame when upscaling (was about 13 px) and from 0.1 % of the width at 1:1 (was 0.5 %). A recorded Silent Hill f turn,
upscaled 1.5x (upscaled / coarse, where trailing shows / steadiness, in dB; a plain stretch 32.68 / 43.94 / 29.78):- DLSS 32.82 / 43.18 / 30.11 -> 34.02 / 47.17 / 31.13
- FSR 33.44 / 45.47 / 30.59 -> 34.10 / 47.84 / 31.20
- XeSS 33.61 / 46.42 / 30.72 -> 34.07 / 47.82 / 31.18
World of Warcraft walking, 1:1 with DLSS: coarse 45.70 -> 47.82. The owner's verdict, live in World of Warcraft: all three "worked and
look better". New "Lean from" slider under "Steady in fast motion" to set it by eye (automatic by default). -
No more flicker while moving at 1:1 with DLSS. Model L pulses on a four-frame cycle when DLSS runs at the screen's size (found
offline on recordings: about 9 times a second at 36 fps); E does not. The DLSS model list has a new default, Auto: E at 1:1 (DLAA),
L when upscaling. A model chosen before is kept (pick Auto once to switch). -
Cost, live at 4K 1:1 in World of Warcraft (RTX 4070 Ti SUPER, GPU time a frame, the motion estimate's share in brackets): DLSS
4.20 ms (1.64), FSR 3.3-3.7 ms (1.1-1.5), XeSS 2.7-2.8 ms (0.7). -
Before / after hotkey: the upscaler addons now show the corner square too (green the upscaled picture, red the original, amber
the split), as Neural Rendering does. -
Neural Rendering: Temporal smoothing works against the model's flicker in motion now: it keeps its history where the frame itself
still matches along the motion, and holds back only where something new came into view. On by default (0.6) for new settings; a saved
"off" is kept, so move the slider to try it. Still the weakest part in motion: more to come. -
Motion by shape (test): a checkbox in the upscalers' panel (off by default) matching the motion by shape and edges rather than
brightness, for scenes where the light changes. -
Tools:
nr_srevalscores steadiness (shimmer from frame to frame) and prints each frame's change from the one before; the motion
estimate's block cost can be varied (meanweight,gradweight). Findings indocs/frame-generation-research.mdand
docs/lsfg-vk-study.md(how Lossless Scaling's frame generation is built, studied from lsfg-vk). -
Known: Neural Rendering still flickers in motion; open items in
handoff/BACKLOG.md. Frame generation of our own is not in this
release.