perseval-BLR/DLSS5-NeuralScreen v1.7.0

v1.7.0 - Boost: up to 80% more frames for the same picture, and the neural calls move to a module of their own

on GitHub

HDR - soon, stay tuned)

Boost

The network is capped at 2560x1440, and until now it ran at the full frame size no matter where the resolution slider stood - the output came back bit-identical at every position of it, measured on four real 4K frames. The slider was decorative.

Boost makes it real. The network runs at the work resolution and its edit is composited back onto the untouched native frame, so text and edges stay 1:1 while the expensive part gets cheaper. On a 5070 Ti at 4K:

At 1:1 on a game scene, code in an editor and photographic content the difference is not visible - the native frame stays the anchor. Without that composite the same setting is visibly soft, which is why it is always on underneath.

A switch directly under DLSS 5, off by default, with the resolution slider under it. All of the above was measured on still frames; how it behaves in motion has not been tested here, and until it has, this is opt-in.

Where the neural calls come from

nvngx_dlssnr.dll decides whether to answer by looking at the module a call returns to: its path has to contain nvngx.dll. It does not look at the process name at all. Measured here in four runs - the call from the executable is refused, the same call from a DLL carrying the substring succeeds and creates the feature, that DLL renamed is refused again, and the substring counts anywhere in the path.

So the calls now leave a small module of our own, nvngx.dll_ns-forwarder.dll. Nothing changes on screen; what changes is that the worker no longer has to be an executable named nvngx.dll - the thing antivirus vendors flag, that cannot be signed sensibly, and that people keep finding in the task manager. Renaming it is the next step.

Fixed

• The capture could open on a different card than the network. NS_GPU is a DXGI index and the network treats it as a wish - an index that is not a usable NVIDIA card falls back to the first one that is. The capture took it literally. On a hybrid laptop adapter 0 is the integrated GPU, and that is the shipped default, so the network ran on the NVIDIA card while the capture sat on the iGPU - and the frame cannot cross adapters. Nothing on screen, no error (#34).
• On a second monitor the picture could stay absent until you switched away and back (#33). One shared auto-reset event served every GPU wait in the worker; a wait woken by someone else's completion was reported as a failure, and the next wait inherited the next stale signal. The pipeline stayed one step behind for good - in one user's log, 2766 consecutive timeouts that ended the moment the pipeline was rebuilt.
• HDR is now reported once, asked of the display actually being captured. The network is trained on SDR; on an HDR desktop the result reads as "too bright, and the sliders do nothing".
• Saved presets never reached the disk. The menu said "Preset saved", the write really did succeed, and the preset was gone on the next launch.
• A card that cannot run the pass stopped being retried forever. The verdict is printed once, on the first frame, and forty later log lines used to bury it.
• A revive brought the long warm-up back on pre-Blackwell cards - the shortening that exists to prevent a restart storm applied to the launch and to nothing else.
• A manual revive left the automatic one armed, so a healthy worker was restarted underneath you up to 30 s later.
• The "this GPU cannot run the neural pass" alert only fired with the menu open - which is the one moment you did not need it.
• The GPU picker showed an empty field when the saved index named no NVIDIA card - on exactly the machines where the setting matters.
• The idle readout no longer disappears after one unrelated log line; the alert font no longer shrinks after a language switch; the window highlight lands on the right monitor; a click on a drop-down's explanation no longer opens the list; Esc closes the drop-down first and the panel second; the screen-capture fallback no longer invents a monitor identity; the Spout2 sender is released on exit; a crash in the output search is gone.

The test suite grew from 67 checks to 92 along the way.

Files

• neuralscreen-v1.7.0-full.zip - everything inside, unpack and run
• README.md / README.ru.md - the guide
• TECHNICAL.md / TECHNICAL.ru.md - how it works

Controls

Num2 menu · Num1 NR on/off · Num3 screenshot · Num0 record · Num4/Num6 resolution · Num5 one window · Ctrl+Alt+Q quit