Nvidia DLSS isn’t magic, and this FSR hack proves it
Nvidia’s Deep Learning Super Sampling (DLSS) has been an undeniable selling point for RTX GPUs since launch, and AMD’s attempts to fight back haven’t exactly been home runs.
But what if FidelityFX Super Resolution (FSR) could provide the huge performance gains of DLSS without all the restrictions imposed by Nvidia? If it sounds too good to be true, I wouldn’t blame you. After all, Nvidia’s special sauce of machine learning wasn’t meant to be easy to replicate.
Well, hold on to your hats because a modder recently discovered how easily FSR could ape DLSS. And after trying out the solution myself, it’s made me more excited about the potential for FSR than ever.
What we have now
Before we get to the mod itself, it’s worth setting the stage for how we got here. FSR was AMD’s first attempt at a DLSS killer, and unfortunately it left a bad taste in our mouths. Despite the rapid introduction of the first generation of FSR 1.0, performance and image quality did not cut it.
All this changed with the release of the technology’s second generation. I have tested FSR 2.0 in the launch title, Deathloop, and the results are clear: DLSS gives a slightly higher performance boost, but FSR 2.0 is almost identical in terms of image quality. Based on Deathloop, you should use DLSS if you can, but FSR 2.0 is a very close second if you don’t have a supported GPU.
My expectations were further exceeded when I tested god of war, sees the margin with DLSS shrink even more. In fact, FSR 2.0 was actually around 4% faster than DLSS with the Ultra Performance preset. You don’t trade much of anything in terms of image quality either. Even with the intense Ultra Performance preset, it’s almost impossible to see any differences between FSR 2.0 and DLSS while playing.
This is the real deal. The only problem? FSR 2.0 is technically available, but it is not seeing the rapid adoption that the first version did. It is available in only four games now: Deathloop, Farming Simulator 22, God of War, and Tiny Tina’s Wonderlands. The upcoming list isn’t that exciting either, with the headline Hitman 3, Eve Online, and the recently delayed Abandoned.
Hence the need for a seemingly impossible solution that takes the goodness of FSR 2.0 and extends its effect to as many titles as possible. And that’s where the fun begins.
A look into the future
About a month ago, modder PotatoOfDoom released an FSR 2.0 “hack” for Cyberpunk 2077. What the modder realized was that DLSS and FSR 2.0 basically require the same information – motion vectors, color values and the depth buffer. It allowed PotatoOfDoom to create a simple instruction translation using the DLSS backbone to send FSR 2.0 instructions. It’s like how Wine works for Windows games on Linux, according to the modder.
I’ll circle back to what these similarities between DLSS and FSR 2.0 mean, but first let’s get games out of the way. I followed the instructions and was able to implement the mod in Cyberpunk 2077, Dying Light 2, and Doom Eternal – all games that do not currently support FSR 2.0. Eternal doom was the only game that struggled with the mod, blocking the DLSS option in the settings menu entirely. It was a no-go.
But Cyberpunk 2077 and Dying Light 2 was an absolute treat. The mod isn’t quite as powerful as a native implementation, but it’s still very close. The difference is less than 10% at most, even with all settings cranked up to 4K (including the highest ray tracing options).
Image quality was just as good, even on this self-penned hack. In a still image, Dying Light 2 actually looked a little better with FSR 2.0 and it was almost identical in Cyberpunk 2077. The main difference, as was the case in god of war and Deathloop, is that FSR 2.0 does not handle distant fine details as well. You can see it on the phone lines in Cyberpunk 2077 under. It’s probably close.
DLSS and FSR 2.0 look mostly the same with a still image, but it’s the movement that matters. I saw heavy ghosts inside Dying Light 2 which was not present with DLSS or FSR 1.0, and flat textures cause some problems with masking.
Certain elements, such as smog from the sewers Cyberpunk 2077 screenshot below, do not include motion vectors. FSR 2.0 and DLSS get around the problem of masking the element (as in Photoshop) so that it is not included in the supersampling. Unfortunately, they perform the masking in different ways, leading to the nasty pixelation with the FSR 2.0 hack that you can see below.
Even with these issues, it’s remarkable how close DLSS and FSR 2.0 are, both on a gameplay and technical level. PotatoOfDoom summed up how much they share in an interview with Eurogamer: “I expected to work with [adding FSR 2.0] for several days, but was pleasantly surprised that it only took me a few hours to integrate.”
The point is not that you should necessarily go out and use this mod to add FSR 2.0 to every game. Rather, this mod reveals the deep similarities between DLSS and FSR 2.0 – something Nvidia might not want to admit.
Take deep learning out of supersampling
DLSS is about machine learning; it’s there in the name. And to this point, Nvidia has insisted for years that DLSS only works on its latest graphics cards because they provide the AI cores necessary to perform the supersampling. That’s true, but FSR 2.0 is proof that the benefit of AI is small and mostly unnecessary.
A big reason why Nvidia’s GPUs sell above list price is DLSS, even though it doesn’t need to.
There are many similarities between DLSS and FSR 2.0, even when it comes to Nvidia’s machine learning bit. DLSS uses a neural network and FSR 2.0 uses an algorithm, but both are fed the same inputs and use the same general system to render the final output. The fact that PotatoOfDoom was able to develop a mod that works across multiple DLSS titles in a matter of hours is a testament to that.
The main problem now is not that DLSS is bad – it’s excellent and you should use it if you can – but that the feature is exclusive to only a few expensive graphics cards. Even as GPU prices drop, Nvidia’s low- and mid-range models continue to sell at above list price. And a big reason why is DLSS, although it doesn’t have to be.
General solutions such as FSR 2.0 and Unreal Engine’s TSR (temporal super resolution) are the way of the future. They work with pretty much all modern hardware, and developers consistently insist that they only take a few hours to get up and running.
DLSS doesn’t have to go away, but it would be nice to see Nvidia leverage its relationships with developers to get a general supersampling feature in games that already support DLSS. And no, Nvidia Image Sharping, which is basically FSR 1.0, doesn’t count.
FSR 2.0 is truly impressive, but gaming support holds it back. Far more games support DLSS than even FSR 1.0, and the official list of four FSR 2.0 is embarrassing. I’m also not excited about too many of the upcoming FSR 2.0 titles, with the list mostly consisting of older or smaller games.
PotatoOfDoom’s mod is a hopeful sign, but we need more FSR 2.0 games for it to even stand a chance against DLSS. It may be tempting to root for AMD here, but it’s important to remember that DLSS still has a slight edge and is supported in far more games. AMD has a lot to cover and FSR 2.0 isn’t being added to games at nearly the same rate as FSR 1.0 was.
Nevertheless, it will be interesting to see how the dynamic between DLSS and FSR 2.0 adapts for the rest of the year. After all, AMD released the FSR 2.0 source code in June. For now, DLSS is still the way to go for its gaming support and a little better image quality, but that’s not the selling point of an Nvidia GPU that it once was.
This article is part of ReSpec – an ongoing bi-weekly column that includes discussions, advice and in-depth reporting on the technology behind PC gaming.
The editors’ recommendations