Neural Rendering is cleaner on black, and runs lighter by default. No glow or green tint beside HUD boxes on black (seen live on the owner's machine: "significantly better"), a new setting to run the model on every 2nd to 4th real frame, and new defaults from the settings the owner ended on. Checked offline and live on one RTX 4070 Ti SUPER; other cards not yet.
- Neural Rendering: a setting "Run the model" (on every frame, every 2nd, 3rd or 4th real frame). The owner's NR runs too heavy on their card, with Auto quality already at its floor: the model's cost is mostly fixed (from their logs about 3 ms a run whether the
model sees 480x272 or more; the 8.5 ms the backlog records at 1912x1080 fits "about 3 ms plus 2.7 ms a megapixel"), so a lower resolution gives little and running it less often gives most. The frames between runs are shown with the last result moved along the motion, as the generated
frames already are. Measured offline on four recordings (nr_nreval, new line "how old a result may be"): the model's change is 4 to 8 levels of 255 big on average; a result two runs old, moved along the motion, is 0.04 to 0.23 levels further from the frame's own result than one run old (0.10 to 1.43 against 0.06 to 1.20), about 1 to 3 % of the change.
Every 2nd frame should take about 1.6 ms back from each real frame at 58 real frames a second. Every 2nd frame by default since this release; Auto quality may still ask for more on top. Host scenarioevery2; settings test. - Neural Rendering: pure black no longer gets a glow or a tint at a small model size. The owner saw a halo around HUD boxes and a green cast on the black background on the character-select screen, less at 0.6 and gone at 1.0. The model's result is stretched over the frame, so beside a sharp edge it spills onto the black side, and on black it leans green. Measured offline on that clip (
nr_nreval piclog=1): black shown as G 0.17 of 255 at 0.25 and 0.04 at 0.6; the strip beside the box 0.87 and 0.47. The compose now gives a pixel at or near pure black (under 0.04 in the picture's own view, about 10 of 255) no positive change (darkguard=0innr_nrevalshows it without); with it both measure 0. Seen live by the owner: "significantly better"; 0.02 was tried first and 0.04 looked right. Spill onto dark but not black neighbours is not covered (that wants an edge-aware stretch of the result). - Neural Rendering: new defaults (the settings the owner ended on after the live tests). Model resolution 1.0 as the ceiling with Auto quality on (time budget 6 ms, lowest resolution 0.6), Run the model every 2nd frame, Temporal smoothing 0.50, Sharpen 0.50. On the owner's RTX 4070 Ti SUPER (WoW at 2560x1440 to 3840x2160, HDR, frame generation x3) the model ran at 0.6 in 5 to 6 ms every 2nd frame with the game's frame time unchanged. Only new installs and settings never saved take them: a saved setting is kept. Colour (gamma, vibrance) stays neutral. A smaller card lowers the resolution by itself, down to 0.6, then runs the model every 3rd and 4th frame.
nr_nrevalprints how old a result may be and the model's GPU time. (The time is only trustworthy under load: on an idle card the clocks stay low and a run takes ten times as long as in a game.)
Zip LSAddonManager-0.9.28-x64.zip, SHA-256 ac1a8b9e58d9299bf3f8f6a7dad7ec3ccfbffeb04d7f070bf1c81443e0163b75.