perseval-BLR/NeuralScreen

NeuralScreen – DLSS‑5 neural up‑scaling for your Windows desktop

What it is – A Windows‑only utility that captures whatever is shown on your screen (the whole desktop or a single window) and runs NVIDIA’s DLSS 5 neural renderer on it in real time. The processed frame is then composited back so you see a sharper, higher‑quality image across games, videos, photos, etc.

How it works – The program ships a small Python‑based runtime that loads the official NVIDIA DLSS 5 DLLs (nvngx_dlssnr.dll for up‑scaling, nvngx_dlssg.dll for frame‑generation). It captures the screen, feeds each frame to the neural network at a reduced internal resolution (the Boost mode), and then overlays the network’s output onto the original frame. The overlay is drawn by a lightweight tray application that can also record or screenshot the processed output.

Key features

• Full‑desktop or per‑window processing, selectable from the on‑screen menu.
• Three built‑in models (Default, Natural, Cinematic) tuned for different visual styles.
• Adjustable strength via Profile sliders and a Boost option that trades a little extra latency for higher frame‑rates.
• Optional DLSS 4.5 Frame Generation (×2/×3/×4) for experimental motion‑interpolation.
• Hot‑key driven UI (Num‑pad keys, configurable) and a simple tray‑menu.
• Built‑in MP4 screen‑recording (with system sound) and screenshot capture.
• Spout2 output for feeding the processed frames into OBS or other capture software.
• Multi‑language UI (12 languages) and HDR‑compatibility toggle (experimental).

System requirements

• Windows 10 or Windows 11 (both reported to work).
• An NVIDIA RTX GPU of Ada Lovelace (RTX 40‑series), Ampere (RTX 30‑series) or older RTX 50‑series. Turing‑based RTX 20 cards start the program but the image is not processed.
• Up‑to‑date NVIDIA graphics driver (the runtime talks directly to the driver).
• No additional software; the release archive contains its own Python interpreter and the required NVIDIA DLLs.

Getting started

1. Download the latest release archive from the GitHub Releases page and unpack it.
2. Run NeuralScreen.exe (or the accompanying NeuralScreen.vbs to bypass the unsigned‑publisher warning).
3. Press Num2 (Num Lock must be on) to open the overlay menu. Use the listed numpad shortcuts to toggle rendering, change resolution, take screenshots, start/stop recording, etc.
4. Choose Fullscreen or Window mode under Source to process the whole desktop or a single window. The overlay follows the selected window automatically.

Known limitations

• True full‑screen exclusive games cannot be overlaid; you must use borderless‑window mode.
• HDR handling is experimental and currently limited to SDR output; recording and Spout exports remain SDR.
• Maximum processing resolution is capped at 2560 × 1440; the final output is always your native monitor resolution.
• Latency of the neural pipeline is about 40–60 ms (≈17 ms with Boost), which is fine for everyday use but not suitable for competitive gaming.
• Rotated displays (90°/270°) are not yet supported; 180° rotation works with a flip.
• Multi‑GPU setups on Windows 10 are still experimental.

License – The project’s own code is released under the MIT license. The bundled NVIDIA runtime DLLs are included unmodified for research/educational use only and remain NVIDIA’s property.

---

NeuralScreen is a genuine AI‑related software project that repurposes NVIDIA’s DLSS 5 neural rendering technology for general‑purpose desktop enhancement.