Affiliate disclosure: This guide contains affiliate links. As an Amazon Associate, we earn from qualifying purchases at no extra cost to you. Learn more.

Every Game the DLSS 5 Mod Has Been Tested On, and How Badly It Fights the Art Style

2026-08-29 · News and analysis

In the days since the DLSS 5 neural rendering library leaked out of an NBA 2K27 build, modders have injected it into roughly 19 games. Some of those results are genuinely striking. Some look like a phone camera beauty filter was applied to a beloved classic, because functionally that is close to what happened.

There is a clear pattern in which is which, and it is not about how good the game looks. It is about how far the game's intended look sits from photorealism.

This is a running list. If you want the background on the leak first, start with our explainer.

The Pattern, Up Front

The model has been trained toward photorealism, and it only touches character models right now, with seven style presets available.

That produces three rough outcomes:

Modern realistic games: the effect is subtle and often reads as an improvement. The game was already aiming at photorealism, so a model pushing it further is aligned with the art direction rather than fighting it.

Older realistic games: the biggest visible change, and the most divisive. Skyrim is the headline case. Faces become dramatically more detailed and dramatically less like the characters they were.

Stylised games of any era: the worst fit. The whole point of a stylised look is that it is not photoreal. A model whose only setting is "more photoreal" has nothing useful to offer here.

Older Games: The Most Dramatic and Most Contested

The Elder Scrolls V: Skyrim (2011). The demo everyone has seen, put together by PureDark inside his existing Skyrim upscaler. Lighting is reworked substantially, and character faces go from 2011 Bethesda to something approaching photoreal. It is technically impressive and it is also the clearest example of the problem. Skyrim's characters look weathered and rough on purpose. Smoothing them into conventionally attractive photoreal faces is not a restoration, it is a replacement. Reported performance was 80 to 120 fps down to around 60 in character creation, and 45 to 50 outdoors.

Grand Theft Auto: San Andreas (2004). The most aggressive mismatch on the list. San Andreas has a hazy, sun bleached, deliberately low fidelity look that is inseparable from what the game is. The mod redoes the lighting and sharpens CJ's face considerably. Whether that is an improvement depends entirely on whether you think a 2004 game should look like a 2026 one.

Grand Theft Auto V (2013). Better fit than San Andreas, because GTA V was already reaching for realism. Still noticeably changes faces.

Control (2019). The first game anyone got it working in, courtesy of RenoDX, and the source of the most quoted benchmark. Deeper shadows and reworked skin shading. Control has a very deliberate brutalist visual identity, and this softens some of it. Cost roughly half the frame rate at 4K.

Assassin's Creed IV: Black Flag Resynced. A remaster receiving another layer of modernisation on top.

Cyberpunk 2077 (2020). One of the better outcomes. The game was built around high fidelity character rendering, so the model has something to work with rather than something to override. Also the cheapest performance cost measured so far, around 25 percent on an RTX 5090.

Recent Games: Subtle and Mostly Fine

These are the titles where the effect is least controversial, because the games were already targeting the same look the model is trained on.

Red Dead Redemption 2
The Last of Us Part II Remastered
Hogwarts Legacy
Starfield
Silent Hill 2 (the remake)
Metal Gear Solid Delta: Snake Eater
Monster Hunter Wilds
Senua's Saga: Hellblade II
Final Fantasy VII Rebirth
Wuchang: Fallen Feathers
Black Myth: Wukong

Several of these, notably Starfield and Hogwarts Legacy, are on NVIDIA's official DLSS 5 launch list for Fall 2026. So the modded version is a preview of something that is coming properly, with developer integration, in a few months.

NBA 2K27 deserves its own mention as the game the library leaked from. It contains no DLSS 5 controls of its own. The files were simply packaged into the early access build by mistake.

The One You Should Not Touch

Counter-Strike. It is on the tested list. It should not be.

This mod works by DLL injection, which is the same technique cheats use and exactly what anti cheat systems are built to catch. Running it in a game with VAC is asking for a ban on an account you probably cannot replace. We go through this properly in the risks article.

Competitive shooters are also the games where this makes the least sense. You would be spending 40 percent of your frame rate on more photorealistic character models in a game where people turn settings down to see enemies more clearly.

What This Tells Us About Older Games And AI

There is a broader point here that outlasts this particular leak.

The retro community has spent 20 years building tools to make old games look right on modern displays. CRT shaders, scanline simulation, aperture grille masks, integer scaling, hardware upscalers like the RetroTINK-6X CE. Every one of those tools is trying to reproduce the output the game was designed for.

Neural rendering is doing the opposite. It is inferring new detail that was never there, toward a target the original artists did not choose.

Neither approach is automatically wrong. But they are not the same activity, and the Skyrim and San Andreas results make the difference unusually easy to see. We wrote that up properly in AI upscaling versus preservation.

For what it is worth, the emulation world has a version of this argument already. Internal resolution upscaling in DuckStation or PCSX2 renders the original geometry at higher resolution, which is a real middle ground. Nothing is invented, the same polygons are just drawn with more pixels.

Frequently Asked Questions

How many games has the DLSS 5 mod been tested in?

Around 19 as of late August 2026, including Control, Cyberpunk 2077, Skyrim, GTA V, GTA San Andreas, Red Dead Redemption 2, Starfield, Hogwarts Legacy and Silent Hill 2. The list has grown quickly since the library leaked and will likely keep growing.

Which game shows the biggest DLSS 5 difference?

Skyrim, by a wide margin. Its 2011 character models change dramatically, with much more detailed and much more photorealistic faces. That is also why it is the most criticised example, since the characters end up looking like different people.

Does DLSS 5 look good in older games?

It depends what you want. It makes them look more photorealistic, which in a stylised or deliberately low fidelity game like GTA San Andreas works directly against the original art direction. In games that were already aiming at realism, such as Cyberpunk 2077, the results are much better received.

Which game did DLSS 5 leak from?

NBA 2K27. The neural rendering library was packaged into an early access build by mistake. The game itself has no DLSS 5 options.

Can I use the DLSS 5 mod in multiplayer games?

No. It uses DLL injection, which anti cheat systems are designed to detect. Counter-Strike has been tested with it, which is a good way to get banned. Single player only.

Are any of these games getting official DLSS 5 support?

Yes. Starfield and Hogwarts Legacy are both on NVIDIA's announced launch list for Fall 2026, alongside Resident Evil Requiem, Assassin's Creed Shadows and The Elder Scrolls IV: Oblivion Remastered.

Why does DLSS 5 only change character faces?

The current modded implementation targets character models specifically, with seven style presets. That is a limitation of how the leaked library has been wired into games it was never built for, rather than a description of what the finished technology will do.