Perseval

можешь все закидывать. Пока с амд туго идет(

Так прошел же давно игрулю) Сделал пост, чтобы сделать перепись мамкиных советчиков по покупке карточек. И усё

https://github.com/perseval-BLR/DLSS5-NeuralScreen/releases
EN:
NeuralScreen v1.5.2:
- 12 UI languages (EN/RU/FR/DE/ES/IT/PT/PL/UK/ZH/JA/KO), proper CJK fonts
- Audio limiter for recordings - no more clipping (peaks up to +7.9 dBFS folded to 1.0)
- Reworked menu: live stats on the main page only, MODE instead of WORK, DLSS 5 ON/OFF toggle, Actions section, footer = Quit only
- Default profile: Natural
- Every control covered by tests

RU:
NeuralScreen v1.5.2:
- 12 языков интерфейса, свои шрифты для китайского/японского/корейского
- Лимитер звука в записи - клиппинг убран (пики до +7.9 dBFS складываются к 1.0)
- Меню переработано: индикаторы только на главной, MODE вместо WORK, тумблер DLSS 5 ON/OFF, секция Actions, в футере только Quit
- Дефолтный профиль Natural
- Каждая кнопка покрыта тестами

https://github.com/perseval-BLR/DLSS5-NeuralScreen/releases/tag/v1.5.1
Hotfix. The v1.4.1+ builds regressed every non-Blackwell card: the architecture spoof target accidentally moved from 0x1B0 to 0x1A0 (arch-table fix), and the leaked runtimes refuse feature 18 below 0x1B0. Same runtime, same cards, one constant - v1.3.0 (spoof 0x1B0) worked on RTX 30/40/50, v1.4.1+ did not.
What is in this release
RTX 30 restored - the universal runtime (310.8.0, sm_75/86/89/120) + spoof 0x1B0, the exact v1.3.0 stack that worked for everyone.
RTX 40 works again - same fix; no more manual DLL swapping.
RTX 50 - unchanged, works.
Honest startup log - [arch] lines now always reach the log and print the real spoof value (the old line hardcoded -> 0x1B0 even while spoofing 0x1A0).
VERSION.txt inside the archive - commit, runtime SHA-256, kernel archs - so you can tell exactly which build you have.
About the v1.5.0 confusion
v1.5.0 was released twice under the same tag: the first zip carried a broken community patch, the second (rebuilt) one carried a Blackwell-only runtime that cannot run on RTX 40 - my mistake, and the archive was not clearly marked. That is on me. v1.5.1 is a single, verifiable archive - the manifest inside states what it contains; no rebuilds under this tag.

What is new in 1.5.0 (https://github.com/perseval-BLR/DLSS5-NeuralScreen/releases)
- RTX 40-series now works. The bundled nvngx_dlssnr.dll is the community re-targeted 310.8.0-RTX40 build (from the RankFTW/rhi-repo mirror): it carries sm_89 (Ada) + sm_120 (Blackwell) kernels, so it runs on RTX 40 and RTX 50. The previous build died on Ada cards with 0xBAD00001 on frame 0 (issues #8, #10, #11, #12). Verified on the bench (5070 Ti): NR ON, ~84 FPS, no artifacts.
- Window picker page. "Select window..." opens the list of open windows; hovering a row outlines the real window on the screen in amber. Only real taskbar windows are listed (no owned/tool/cloaked helpers).
- Visible window-mode exit. The footer "Fullscreen" button returns the capture to the whole screen; in fullscreen mode it shows an "already active" alert. The capture mode ("Fullscreen" / "Window mode") is shown under the GPU line.
- Spout2 bridge (NS_SPOUT=1). The worker can publish its output as a Spout2 shared texture - OBS with the Spout2 Capture plugin records the NR picture directly. Off by default.
- DRED device-removed diagnostics. When the GPU device dies (the Win10 TDR from issue #1), the log contains the exact removal reason.
- Recording at 30 fps instead of 60. Halves the recording FPS cost, the per-frame picture quality is identical.
- Safe passthrough. When feature 18 cannot be created, the worker stays alive and shows the raw frame instead of exiting.

печаль. А мерцания - да - вопрос реализации и адаптации. Без честных векторов движения будет периодически где-то фликерить

https://www.tweaktown.com/news/113445/you-can-now-apply-dlss-5-neural-rendering-to-your-entire-windows-desktop/index.html

https://www.tweaktown.com/news/113445/you-can-now-apply-dlss-5-neural-rendering-to-your-entire-windows-desktop/index.html

Это обход вообще всех оверлеев и работает строго поверх. По этим моментам - скорее всего нужна будет интеграция с каким-нибудь lossless scaling - покопаю в эту сторону - но вероятно непросто будет

По поводу нескольких ускорителей - тут вопрос проверки такого решения) вторая 5070ti у меня есть - но вот поставить физически в доску нельзя будет)