Today, we're posting our full review of the recently released Nvidia DLSS 4.5 - the brand-new Preset M, designed to offer increased upscaling quality in performance mode. An evolution of its existing transformer-based machine learning technology, DLSS 4.5 super resolution works on any RTX GPU currently available, though its use of FP8 acceleration means it runs fastest on GPUs based on the RTX 40-series and 50-series architectures.

An image quality based review typically works best in the medium of video, so I'd recommend you watch the discussion between Alex and John above, but in terms of the need-to-know information, there are a few takeaways here before we into a touch more depth:

  • DLSS Preset K is the original transformer model, Preset M is the DLSS 4.5-orientated upgrade designed for performance mode - typically delivering a 2x2 upscale. So, a native 960x540 becomes 1080p, 1280x720 becomes 1440p and 1920x1080 upscales to a 4K output.

  • There's also a new, computationally more expensive DLSS Preset L, designed for 3x3 upscaling: 1280x720 becomes 4K, for example. We're not covering that today.

  • Using Nvidia GPUs of equivalent performance across three generations - we noted higher hits to performance using Preset M over Preset K when benchmarking Cyberpunk 2077 at 4K resolution.

  • RTX 3070 saw an 18.4 percent drop to performance, RTX 4060 Ti a 6.6 percent drop, while RTX 5060 was impacted by just 4.8 percent. The newer your GPU, the more performant the new DLSS is.

  • There are good improvements to image quality from Preset M in rasterised titles, addressing many (but not all) of our problem scenarios with the existing Preset K.

  • However, there are issues with certain titles using ray tracing, some of which are very similar indeed to the problems we've seen with certain PSSR implementations on the PlayStation 5 Pro.

  • If a game supports DLSS ray reconstruction, that's an entirely different form of DLSS that has not been updated. If you force on Preset M via the Nvidia app then turn on ray reconstruction in a supported game, your override will be ignored.

Let's kick off with the good stuff - while DLSS 4.0 was exceptionally impressive, it was not quite the finished article. For example, we spotted issues with volumetric fog effects, where ghosting and shadows could cause issues. Assassin's Creed Shadows was a good example of this. Additionally, pixelation issues on disocclusion (where new visual information is revealed), were also a slight problem. Aloy's motions in Horizon Forbidden West showcase that problem. In both of these issues, DLSS 4.5 Preset M is a very good improvement.

More neutral are a couple more cases where 4.0 Preset K had issues: Forza Horizon 5 is arted around classic MSAA anti-aliasing, so thin sub-pixel elements like power lines rendered nicely there - an area where all DLSS has struggled over time. This is improved in Preset M over the lacklustre, ghosting results from Preset K - but MSAA still looks better. Similarly, DLSS's impacts on rain rendering in Death Stranding see improvement with Preset M, but the little splashes on Sam's backpack are not improved.

Still, the more you look, the more improvement you'll see - vegetation fizzle in Kingdom Come: Deliverance 2 and the Outer Worlds 2 is improved, anti-aliasing effects on distant lines are much better in Marvel's Spider-Man: Miles Morales, while disappearing particles in The Outer Worlds 2 see fewer of them erroneously disappearing with the new model.

Everything looks good except for the problems we're experiencing with Preset M in a range of games where RT is supported. While some areas - like reflections in The Outer Worlds 2 - remain similar, there are regressions. Control: Ultimate Edition has less stable reflections on rougher surfaces, with flicker and noise. In Star Wars Jedi: Survivor, ambient occlusion on vegetation is less stable and noisier - bringing to mind PSSR on PS5 Pro. That said, here, reflections are more stable on water. That said, Silent Hill 2 has less stable water reflections compared to the older Preset K, but there are similar ambient occlusion issues on foliage.

There's more nuance to the discussion but the bottom line is this. We'd say that if you own an RTX 40-series or 50-series card, switching from DLSS 4.0 Preset K to the new Preset M is generally a good idea in games without ray tracing. Users of RTX 20/30-series cards can use the technology, but arguably the hit to performance makes it less worthwhile. Meanwhile, for RT-based titles, the jury is still out on Preset M. We'd recommend using the Nvidia App to override whatever model is in play and to decide for yourself what's best - but it's clear that there is work to be done here.

And meanwhile, we're very curious to see when and how Nvidia will improve on the existing transformer-based DLSS ray reconstruction preset, or whether they'll save that for the inevitable DLSS 5.0.