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.
Important
Radeon testers wanted. The AMD build lives in a separate repository, NeuralScreen-AMD, and it has not run on a real Radeon yet. Start from native/AMD.md there, and if it does not come up, attach the diagnostic package (Settings -> Program -> Create diagnostic package).
A picture-quality release: steadier NR, far fewer false scene resets, no "jelly" on still backgrounds with NVOFA, and the open tracker bugs. Each change was measured on the worker with known motion before it went in.
Picture
- The neural edit is steadied over time (Boost). The network's output moves a little on every frame even where the picture does not, and motion errors come on top. Only the network's edit is now carried along the motion and blended where a per-pixel test trusts it - the picture underneath is never delayed. Frame-to-frame instability of the edit about halved (pan -43%, moving object -45%, still content -49%), the strength of the effect kept within 2%, no ghosting with wrong vectors, 0.1-0.3 ms a frame.
NS_STAB=0turns it off. - NVOFA "jelly" on still parts of the picture (#151, #141, #95). NVOFA's coarse field spread a moving object's vectors onto the still background around it. A vector is now kept only where it beats "no motion": still background with motion around a moving object 91.7% -> 0%.
- Scene cuts are judged on the picture. The old fixed threshold reset the network's history on every frame of a fast cel-shaded pan and missed cuts between dark or text pictures. The frames are now aligned (shift and contrast) before the decision: false resets 4.8% -> 0.6%, missed cuts 39% -> 1.3%.
- A still desktop no longer shimmers or keeps the GPU busy. An unchanged capture is evaluated a few times, then the last result stays on screen.
- Extra NR passes no longer darken the picture (0.988 -> 0.970 brightness over three passes; now 0.988 at any count).
Fixes
- Window mode is no longer held at 60 FPS (#155): window capture now asks for a 4 ms update interval - 60 -> 120 fresh frames a second on a 120 Hz monitor, 49 -> 63 NR FPS on the reporter's machine.
- A controller that connects and drops again no longer closes the program at startup (#152).
- A capture that keeps switching between 8-bit and 16-bit frames is pinned after the first switch (#151): one driver did it ~3 times a second on 8-bit monitors, each switch costing a frame.
- A runtime that answers success but changes nothing now fails the startup check with a clear message instead of running NR that does nothing.
- The "adaptive exposure" is gone: the runtime never read that parameter (byte-identical output at 0.3, 1.0 and 3.0).
For #142
Two opt-in experiments for an uncapped game in focus under hardware GPU scheduling: NS_GPU_PRIORITY=realtime (needs NeuralScreen started as administrator) and NS_GPU_QUEUE_PRIORITY=high|realtime. Off by default; the log says what each one got. With HAGS off the problem does not occur.
Tests
Suite on the release branch: 290 pass, 1 fail (the runtime manifest, re-pinned by this release), 7 skipped, GUI end-to-end cycle green; after the version commit the static gate and the release-contract tests pass. HDR tests were run on an HDR display; the HDR recording test stays borderline as before (92-107 frames against 96 in 2 s).