perseval-BLR/NeuralScreen v2.6.5

v2.6.5 - FG Lite, a multiplier that follows the display, and a package that can explain itself

on GitHub

Notice. Not affiliated with NVIDIA; NVIDIA, DLSS and the NVIDIA logo are NVIDIA Corporation's trademarks. The bundled NVIDIA runtimes are NVIDIA's property, included unmodified as received, research/educational use only, no warranty, use at your own risk.

This release adds two Frame Generation options from the October scan, and fixes four things a report or a panel showed wrongly.

New

FG Lite (off by default). Network -> "FG Lite" asks DLSS Frame Generation for its preset C (DLSSG.LiteCandidateBlendingEnabled, in the bundled 310.9.16 runtime): cheaper generated frames, which NVIDIA quotes at 10% less FG cost at x2 and 36% less at x6 on RTX 40/50. A change recreates FG, like a new multiplier. Bench, RTX 5070 Ti, 4K: FG's GPU time per real frame 1.99 -> 1.14 ms at x2, 4.45 -> 1.88 ms at x4, 6.93 -> 2.73 ms at x6.

FG matched to the display (off by default). Network -> "FG matched to the display" picks the largest multiplier, up to the one you chose, whose output still fits the screen's refresh rate: a 30 fps video on 144 Hz runs x4, a 60 fps game x2, a 24 fps film x6. A new step is taken only after it has held for two seconds, because a step change recreates FG and a source wobbling at a boundary must not flip it back and forth. The FG row says which step runs. The source's rate is the worker's own count of new pictures; a processed-frame count read a 36 fps source as 37-40. Bench, 143.9 Hz: a 36 fps source settles on x4 within ten seconds, a 72 fps one stays on x2.

Fixes

The log says whether NR was on (#194). The [env] switches: line read a config key nothing writes any more, so it printed "NR on" in every session. A reporter's package said nr_on: false in its settings and "NR on" in its header, three lines above "NR starts off, as it was left". It reads the real switch now, so a report can be trusted on the one thing it is asked about most.

The panel opens the folder the packages land in (#196). Settings -> Program has "Open diagnostic package folder" under Create diagnostic package. The folder is created first, since a fresh install has none until the first package, and a refused shell is an alert rather than a crash.

Skin structure names its own off (#193). Its range is -1.0 to 2.5 and -1.0 is off, but the value cell formatted the number for every position. A reporter set the slider to 0, read "0.00" and reported the face filter still working - which it was, 0.0 is a real middle value. The cell now reads "Off" in the interface language below zero.

Two settings a report cannot do without (#194, #193). fg_lite and fg_match_display travel in the diagnostic package. Without them a package cannot say whether FG was running a cheaper preset or following the refresh rate - which is exactly the difference between a 180 fps cap and a 290 fps reading in the same session.

Tests

• 337 checks passed, 6 skipped (the HDR-only tests - this bench has no HDR display), 2 failures, neither a failure of the shipped code.
• worker: fresh, with the hook (the static gate) compares the worker binary's timestamp with its sources. The branch switch touched those sources after the worker was built; the check passed after native\build-host.bat, which the release build runs anyway.
• test_dda_format_pin.py failed once inside the full run on a busy desktop and passes standalone (160 frames, one injected format change, one rebuild).
• The GUI cycle (launch -> record -> exit) passed, and all 14 GUI-E2E tests passed, including the one-window switch and the menu across it.