Microsoft's DirectSR aims to unify DLSS, FSR and XeSS upscaling technologies

Daniel Sims

Posts: 1,370   +43
Why it matters: AI-based image upscaling techniques from Nvidia, AMD, and Intel have become increasingly useful for improving game performance and image quality. However, competition between GPU manufacturers has made their support across titles spotty. Microsoft plans to reveal its response to the situation at GDC next month.

Update (Feb 29): Microsoft has provided additional details on DirectSR's nature and purpose, describing it as an API that provides game developers with a unified codebase for implementing multiple super-resolution methods, including DLSS, FSR, and XeSS.

Microsoft's solution will likely cause more games to support all three. DirectSR will soon be available in the Agility SDK for testing and feedback.

Microsoft will reveal DirectSR during its DirectX State of the Union presentation at GDC 2024 next month. Details on the technology are scant, but it will likely be Redmond's response to AI-assisted image upscaling technologies currently offered by graphics card manufacturers.

The company's announcement describes DirectSR – presumably short for Direct Super Resolution – as a new DirectX tool that makes it easier to implement image upscaling across Windows devices. While Microsoft hasn't confirmed the exact nature of DirectSR, its stated design goal could help developers and users currently juggling at least three upscaling technologies.

Nvidia's DLSS was the first on the market, and comparisons between super-resolution methods often rank it best in image quality, but it requires the tensor cores only found in the company's recent GPUs. Intel's XeSS is similar, needing hardware from the company's Arc GPUs for full effect, but a less efficient version using DP4a functionality commonly found in other cards is also available. AMD's FSR is software-based and thus mostly hardware-agnostic, but it can appear less stable in motion.

Relatively few games support all three options, frustrating users. Applying the work from implementing one super-resolution technology to the others is simple enough that modders can quickly fill in the gaps when a title excludes DLSS or FSR.

Other system-wide options also exist for games that don't natively support upscaling. The Nvidia control panel (and some titles, like Call of Duty) includes a basic spatial technique called Nvidia Image Scaling that works on any graphics card. The paid version of Lossless Scaling, available on Steam, contains a universal upscaling function utilizing machine learning. Furthermore, Microsoft teased an OS-wide alternative in a recent Windows insider build.

Beta testers can activate Automatic Super Resolution on a per-app basis in the display settings. However, Microsoft hasn't confirmed anything about the feature, which could be related to DirectSR.

If DirectSR isn't simply Microsoft's take on FSR or DLSS, it might be a method of reconciling the numerous upscaling technologies. Engineers from Nvidia and AMD will participate in Microsoft's GDC presentation on March 21, but it isn't clear whether they will jointly speak about DirectSR. A toolchain allowing developers to easily support super-resolution options that take full advantage of every GPU's capabilities would be greatly appreciated if that is indeed what Redmond is working on.

Permalink to story.

Gotta love Microsoft logic hard at work here: There's too much fragmentation for ML upscaling solutions! We're going to just offer a DirectX or even OS level solution instead!

You know why? Because nobody but us is allowed to intentionally fragment the PC market and by the way, this is of course going to be exclusive to Windows 11 because you people have not been upgrading to 11 fast enough: it's almost time to charge all of you all over again for 12 and we'd like to be at least 70% install base before that happens so, chop chop PC guys: upgrade to 11 to get a 'Universal' AI upscaling solution!
If MS is able to give us something like XESS, a unified upscaler that will take a shader path for non-tensor cards and a matrix path for Tensor cards that would be fantastic.

Hell if something like that could completely replace TAA (since we're stuck with it as part of the rendering pipeline in modern engines anyway) since its a core function of DX as opposed to an assortment of side-saddle techniques that would be amazing.
I wonder if this will be able to leverage the AI cores in the forthcoming cpus in the future. I’m interested to see what this looks like comparing an RTX card vs an AMD card and whether this becomes a wrapper for DLSS or FSR and/or an upgrade/downgrade for either card flavour as a result.

Edit****. Ok so it’s a single set of instructions that will leverage/convert to understandable operations on Intel, AMD and Nvidia cards using whatever upscaling hardware/software that manufacturer is using, making it easier to enable all 3 current technologies without having to implement all three separately.
Predicted as much; really no different then every other time dueling NVIDIA/AMD(ATI) features got unified in a future DX revision (Pixel Shaders being a prime example). I called this about a month ago.