Nvidia’s DLSS 5 is the most powerful neural renderer the company has ever shipped, and it’s locked inside a handful of games. A free app called NeuralScreen just broke it out. It spreads that AI upscaling magic across your entire Windows desktop, and the results with old videos are honestly kind of wild.
What NeuralScreen actually is
NeuralScreen is a free, open-source Windows 11 app that applies the DLSS 5 neural renderer to any window on your screen. Games, video players, photo viewers, anything visible. It processes what’s on screen in real time, with intensity controls and a before-and-after slider so you can see exactly what the AI is doing. The project lives on GitHub, built by the modder behind several DLSS 5 experiments.
It comes from the same modding community that spent months squeezing DLSS 5 onto graphics cards it was never meant for. But instead of improving one game, NeuralScreen treats your whole desktop as the canvas.
What you need to run it
The requirements are strict, so check before you get excited:
- An Nvidia RTX 30, 40, or 50 series graphics card. No AMD, no laptop integrated graphics.
- Windows 11.
- A willingness to run software that depends on a leaked Nvidia DLL, because that’s what this is built on.
That last point deserves a pause. NeuralScreen intercepts an Nvidia API call inside its own process and reports a Blackwell architecture to the DLSS 5 runtime, which officially refuses to enable its neural features on anything older. Same trick the modders used, just applied to all of Windows instead of one game. It works, but it’s community software riding on Nvidia’s unreleased plumbing.
First time hearing about DLSS 5?
Quick catch-up, because the naming gets confusing. DLSS is Nvidia’s AI rendering tech, and it has quietly become one of the biggest reasons to buy an RTX card. Older versions upscaled lower-resolution frames or generated extra ones. DLSS 5 went further: it’s a generative neural renderer that redraws detail the GPU never actually computed, which is why it looks startlingly good and runs so heavy.
Nvidia launched it officially on RTX 50 series cards, with a big game release as the showcase, and RTX 40 support promised later. That gatekeeping is exactly what the modding scene has been picking at ever since, and NeuralScreen is their boldest crack at it yet.
The best use case isn’t gaming
Here’s the part nobody expected. DLSS 5 is Nvidia’s heaviest model, and experimental builds cut frame rates roughly in half even on new cards. For gaming, that’s a brutal trade. For video, it barely matters.
Old videos are where NeuralScreen shines. That 480p clip from 2009, the compressed family footage, the low-res concert recording: DLSS 5 redraws them with plausible sharpness while they play. RTX cards could already upscale video in some browsers and players, but NeuralScreen works over practically any window, no importing files into a dedicated enhancer, no re-encoding, no waiting.
There’s a built-in screenshot tool and video recorder too, so you can capture the enhanced result directly from the app.
One honest warning before you try it. DLSS 5 generates detail, it doesn’t recover it. The AI invents plausible texture based on what it thinks should be there. For making an old video watchable, great. For forensic accuracy, no. It’s enhancement, not restoration.
The limits you should know
NeuralScreen has real constraints, and the GitHub page doesn’t hide them:
| Limit | Detail |
|---|---|
| Resolution cap | Neural pass tops out at 2560 x 1440 |
| Latency | Adds roughly 40 to 60ms, noticeable in fast games |
| HDR | Not supported |
| Fullscreen games | Exclusive fullscreen can’t be overlaid, use borderless |
| Performance cost | DLSS 5 can halve frame rates on newer cards |
| GPU support | RTX 30/40/50 only, and 30 series runs it worst |
One limit deserves its own paragraph. The project’s own documentation warns that running it over online competitive games could trip anti-cheat software, because of how the overlay hooks into the rendering pipeline. Don’t risk your ranked account to watch Valorant look prettier.
On an RTX 30 series card, expect the neural pass to run, but not necessarily run well. Reports from early users put it at single-digit to 20 frames per second at 1440p on older cards. On a 40 or 50 series card, video enhancement should feel smooth since the latency matters far less when you’re watching instead of aiming.
How to set it up in 5 minutes
- Grab the latest release from the NeuralScreen GitHub page and unzip it. The archive bundles everything, including the 165 MB neural DLL that GitHub wouldn’t host in the repo itself.
- Run the app. Windows will warn about an unknown publisher since it’s unsigned community software.
- Open any old low-resolution video in your normal player.
- Toggle NeuralScreen on and drag the intensity slider until it looks right, using the before-and-after screen wipe to compare.
- If you like the result, use the built-in recorder to capture the enhanced version.
That’s it. No encoding queue, no cloud upload, no subscription. Your GPU does all the work locally.
When it looks wrong
Sometimes the output gets weird. Faces in old footage can drift into uncanny territory, text overlays turn to mush, and heavy compression artifacts can confuse the model into inventing details nobody filmed. When that happens, don’t delete the app. Lower the intensity slider first, then drop the resolution the neural network runs at, which is the main speed and quality lever. If a video still looks worse with the filter on, it just isn’t a good candidate. Not every source clip deserves the treatment.
Why this matters beyond the novelty
Two bigger stories are hiding in this app.
First, it’s another example of the modding community dragging AI features across hardware lines that companies lock down. Nvidia wants DLSS 5 selling RTX 50 cards. Modders think a 3080 owner paid for neural hardware and should get to use it. Both sides have a point, and apps like NeuralScreen are how that fight plays out in practice.
Second, it previews where desktop AI is going. Nvidia’s own research already turns idle GPUs into personal AI systems, and we covered how Nvidia PAIR does exactly that. Real-time neural processing of everything on your screen, not just games, is the obvious next step. NeuralScreen just got there first, on a leaked runtime, for free.
And honestly, the fact that the developer reportedly built parts of it with Claude’s help tells you where solo software is heading too. One person plus an AI coding assistant just embarrassed a trillion-dollar company’s product roadmap. We’ve seen this movie with Nvidia’s AI PC plans, and it keeps repeating.
If you run AI locally at all, the idea that your GPU can redraw your entire desktop in real time is a glimpse of where Windows is headed. Microsoft wants AI in every corner of the OS. The community is already shipping it.
Should you install it?
If you have an RTX 40 or 50 series card and a pile of old videos or compressed screenshots you wish looked better, yes. It’s free, it runs locally, and the before-and-after slider sells itself. Just treat it as experimental: a driver update could break it any week, and the leaked-DLL foundation means Nvidia could shut the whole thing down with one runtime change.
If you’re on older hardware, skip it. The performance hit isn’t worth blurry results on unsupported cards.
Takeaway
Grab NeuralScreen from GitHub, run one old video through it, and see the difference yourself. It costs nothing but a few minutes, and it’s the cheapest look at the AI-powered desktop you’ll get this year.
