There's more to DLSS 4.5 than we've previously talked about. We've already covered what Nvidia refers to as "Preset M" - a new, transformer-based algorithm for performance mode (effectively 2x2 upscaling), but two models were included in the package. "Preset L" is the most advanced DLSS model Nvidia has ever produced, looking to deliver excellent quality with a far more ambitious 3x3 upscale. With 4K resolution, for example, Preset L is upscaling from just a base 720p image. Drop down to 1440p and - remarkably - Preset L is reconstructing from just 480p. Results with Preset M were mixed, so how does Preset L stack up?

As usual for content about image reconstruction technologies, watching the video is the best way to go in getting an idea of how good the quality actually is, but here are a few bullet points to kick off with:

  • Similar to DLSS 4.5 Preset M, games don't native support it. You need to use the Nvidia app to override the pre-existing settings in DLSS-supported games.

  • This is the most computationally expensive DLSS yet. On an RTX 3070 at 4K resolution, upscaling from 1080p, Preset M is already 18 percent slower than the original transformer model (Preset K). Preset L is even heavier. It is 27 percent slower. The performance gaps will be lower on RTX 40-series and 50-series cards.

  • However, it's understandable why Nvidia recommends ultra performance mode for Preset L. It is faster than performance mode with either of the previous transformer models. So it's a question of whether the quality of ultra performance with Preset L is better than the other options in performance mode.

  • In terms of challenging trouble spots for DLSS 4.5 Preset M, we note the same issues with particle reconstruction and wire rendering on Preset L.

  • Aliasing and foliage fizzle we saw in Silent Hill 2 and Horizon Forbidden West are definitely impressed with Preset L over Preset M.

  • "Boiling" artefacts on ray tracing effects seen in Preset M also present with Preset L.

1
An outline of how DLSS costs compare on an RTX 30-series GPU. Preset E is the old CNN model, Preset K is the original DLSS 4.0 transformer, while M and L represent the new DLSS 4.5 offerings.

On the face of it then, in like-for-like testing upscaling 1080p to 4K, there are only modest improvements in addressing the issues we first noted with Preset M. The ray tracing situation in particular is probably the main element that should be addressed. However, there's an interesting wrinkle here. It seems that DLSS 4.5 in both presets does not interact particularly well with denoised RT effects - but a remarkable thing happens if you actually disable the games' denoisers. The problem disappears.

On the one hand, this means DLSS is a better denoiser than some games' standard denoisers - which is kind of funny - but on the other hand, it means DLSS preset M and L have issues when overlayed after stock denoising. That is problematic as we want image reconstruction to be as plug and play as possible without these issues and without having to effectively mod games, but also, we probably want titles to implement things like DLSS Ray Reconstruction anyway.

So far, we've been looking at like-for-like testing against prior DLSS versions with 2x2 upscaling at 4K - but remember, Nvidia's recommended use-case scenario is 3x3 upscaling instead via ultra performance mode. The image quality characteristics are much the same, however, with differences lying in the amount of detail resolved and how competent anti-aliasing is at any given moment.

  • Preset M at 4K performance mode is quite comparable to Preset L in ultra performance mode, despite the massive jump from 4x scaling to 9x scaling.

  • However, Preset L's inner surface detail can be flatter, with reduced normal mapped texture detail.

  • On a general level, Preset L is softer in ultra performance mode than it is in performance mode, so it looks softer. But this is a general trend across many games: the lower the DLSS preset and the more aggressive the scaling level, the softer the image gets.

  • Regardless, Preset L in ultra performance mode can look pretty great - the issues with RT effects is much more of a problem than any drawbacks in using a lower resolution.

    • 2
    DLSS 4.5's issues with RT seems to be down to poor interactions with game denoisers. Here in The Outer Worlds 2, we've completely disabled the game's denoiser and just let DLSS take over. The result is a complete fix to the issue. Click on the image for a closer look and compare the area within the white highlight.

Potentially, Preset L could benefit users of older RTX GPUs if quality holds up in ultra performance mode at 1440p. It's a tough ask bearing in mind the core 480p (!) resolution. Of course, the RT issue is far more pronounced here, while sub-pixel detail is significantly compromised at lower resolutions giving DLSS far less to work with, resulting in ghosting artefacts.

However, perhaps the biggest issue with ultra performance mode at 1440p - even with Preset L's high quality upscaling - is that most games are just not designed around this scenario and have a surprising amount of oversights or compatibility issues. For example, Ratchet and Clank: Rift Apart's RT ambient occlusion completely breaks at this scaling level, while strange lines can follow the camera around in Remedy's Control. Forza Horizon 5 looks fine until water splashes onto your car, causing flashing speckles, while tessellated deforming surfaces cause some remarkable artefacts. Ideally, this would not happen if every aspect of the engine's parameters were set up to scale with output vs input resolution, but that cannot be expected across the board for all games.

Ultimately, Preset L has advantages over Preset M, but it shares the same issues with ray tracing noise, which - curiously - appears to be caused by some interaction with games' standard denoisers being on at the input resolution. DLSS 4.5 produces superior results when they are turned off. Preset L is nice for a 4K resolution output at ultra performance mode on modern RTX GPUs, but becomes less impressive at lower resolutions like 1440p. Still, there's nothing wrong in giving it a go yourself, but my recommendation would be to use an existing, cheaper model like Preset K or Preset M at a higher internal resolution.