Summary
The release of DLSS laid the foundation for what was to come, and while initial reactions were mixed, it did prove one thing — upscaling was the future. Now, while that’s not necessarily always a good thing these days, there’s no denying the usefulness of this kind of technology. DLSS continues to remain the best upscaler in PC gaming, with Nvidia constantly adding newer features to it. Each DLSS version brought with it a plethora of changes and new, adjacent features that still share the same generic name. And herein lies the problem. DLSS is used to refer to way too many things, and untangling that mess is a nightmare. All the DLSS, explained There are far too many to keep track of When it comes to DLSS, we can split it into two trees: traditional upscalers and DLSS-adjacent technologies. For the former, it’s as simple as going from DLSS 1 through DLSS 3.5. For the latter, though, it’s where things start getting a bit messy. These include things like DLAA and DLSS 5, which are completely different kinds of technologies. Confusing naming scheme aside, here’s the rundown for the traditional DLSS upscaler:
- DLSS (original model): The first version of DLSS was trained on a per-game basis, requiring Nvidia to optimize each game. This was also the worst version of DLSS, having noticeable smearing and blurry visuals.
- DLSS 2: The newer, updated version of DLSS was released in 2020 and was a major step up. It used a generalized network trained on a variety of data (instead of being restricted to per-game optimization) models.
- DLSS 3: Introduced with the RTX 4000 series, DLSS 3 improved on DLSS 2 by adding frame generation. This required specialized hardware, which meant it didn’t work with RTX 3000 and older GPUs.
- DLSS 3.5: A minor update to DLSS 3 that introduced Ray Reconstruction.
- DLSS 4: DLSS 4 was a major shift, swapping to a transformer-based model. It reportedly uses a lot less VRAM and is expected to be faster as well. Nvidia thankfully chose to release DLAA, Ray Reconstruction, and the updated Super Resolution tech for all RTX GPUs, which means the newer model can even be used on RTX 2000-class cards (with caveats).
- DLSS 5: The newest iteration of DLSS. Instead of working as a traditional image upscaler, it takes a completely different approach, choosing to act as an “image filter” of sorts, adding in extra detail and “remastering” games to have a more realistic look. Keeping up? Don’t worry, there’s still more; the DLSS-adjacent technologies, which include the following:
- DLSS Super Resolution: What DLSS is mostly used for, as an upscaler. It renders games at a lower resolution and upscales them to a higher-resolution image, boosting the frame rate in the process.
- DLAA: Also known as Deep Learning Anti-Aliasing, DLAA works using the same neural network as DLSS, but sets itself to the native display resolution instead. DLAA is mainly used for antialiasing purposes — which it does quite well.
- DLSS Ray Reconstruction: An AI denoiser for ray-traced effects and path tracing. Significantly makes the process a lot easier and effectively removes shimmering and smearing.
- DLSS Frame Generation: New tech, generates “fake frames” using an AI model to overlay over the incoming image. Provides the illusion of more FPS, often at the cost of increased latency.
- DLSS Override: Used via the Nvidia app, and lets you inject DLSS into older games. You can also use it to force an updated version of the upscaler in games that shipped with an outdated version. As you can see, DLSS is an umbrella term that is used to refer to way too many things. Nomenclature shenanigans aside, when we talk about DLSS, it’s usually the upscaler only, and nothing else. DLSS should only be used to refer to the default, basic upscaler No fancy frame generation shenanigans DLSS is a broad term at this point, and Nvidia loves to use it interchangeably. Modern DLSS iterations almost always take into account newer “features” like frame generation, which is technically cheating, if you ask me. The frame rate number obtained using frame generation is not indicative of real-world performance. Higher frame gen templates inevitably introduce ghosting and latency, to the point where I would recommend disabling the feature entirely. Marketing and deceptive naming schemes aside, when we refer to DLSS, it should be used to speak for the default upscaler tech. All things considered, even if we take away the frame gen and ray tracing tech, DLSS on its own is still a fantastic upscaler and one that continues to set the benchmark. Unless we count something like DLSS 5, which has no reason to exist. Like it or not, DLSS is still king DLSS kind of pioneered the whole upscaling method and remains the undisputed king. Yes, there have been alternatives such as FSR and the criminally underappreciated XeSS, but they do not get close to DLSS’s level of quality preservation — especially at lower resolutions. In other words, DLSS is still the benchmark. That said, Intel, AMD, and even Apple have made significant strides in developing their own alternative upscalers, with AMD’s FSR 4.0 being a noticeable step up in visual quality. And it continues to impress, even if I’m not a fan of gating modern “features” behind newer hardware. With modern iterations like DLSS 5 choosing to add visual filters to games instead of upscaling them, I’m beginning to think we might have already reached the peak of DLSS. Which is both good and bad. At least it’ll still remain a boon for lower-end devices, at least until the illusion shatters.