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.
Warning
NeuralScreen requires the latest NVIDIA driver. Operation with older drivers or unsupported/non-standard configurations is not guaranteed.
A moving face's shadows no longer trail, the fisheye grows into a camera, recordings show the pointer, and switching windows no longer restarts everything.
Picture
- Shadows on a moving face no longer trail behind it (#160). From 2.2.0 on, soft shading on a slowly moving face lagged for a fraction of a second. Two causes, both changed: NVOFA's zero-motion test took the vectors away from shading that changes by less than a grey level a pixel, and the edit stabilizer kept blending the old shadow where its test could not tell soft content apart. A real change in the edit now passes straight through, and the stabilizer's memory follows the time between frames - a 30 fps game no longer keeps an old edit twice as long as a 60 fps one. Measured with a synthetic soft face at 30 fps; the stabilizer's gains on pans and moving objects are unchanged.
New
- Camera (#159). The main-page switch is now Camera; under it the Fisheye (on by default, so a 2.3.0 setup looks the same) and four amounts, each 0 = off: Webcam noise - now also over an unbent picture -, Vignette, Chromatic aberration and Small sensor (crushed shadows, clipped highlights, less colour, softer corners). Drawn after the network and Frame Generation; in recordings, Spout and screenshots.
- Recordings show the pointer (#131) - the system pointer's own shape, where it was. Not drawn when a game hides it; HDR10 recordings go without it for now.
- Picking another window keeps the worker (#148, #157). From one window to another on the same monitor, with NR on and nothing recording, the switch is a live reconfiguration (under a second) instead of a restart (17-38 s on one reporter's machine).
- Pop-up notifications: All / Errors / None (Settings -> Program). Hidden ones still go to the log.
Fixes
- A lost capture no longer spins a CPU core (#157). In an RDP session (and while UAC or the lock screen is up) the frames came back-to-back, ~500 a second at 92% of a core; they now come no faster than a 30 fps capture's until the capture is back.
Tests
Full suite on the release branch: 307 pass, 5 fail, 7 skipped. The five: the runtime manifest (re-pinned by this release); two checks fixed after the run (the new notification test, and a test stub that lacked the new method; both pass now); test_fg_waitable_drain (borderline, 2.46 ms against ~2.5 - passes on rerun, the FG presenter is untouched); and the pointer test, which missed the pointer twice early on and then passed 10 runs in a row. HDR tests skipped (HDR off).