TraverseResearchWork with us
Menu

Traverse Research

Neural reconstruction and upscaling

For silicon and driver teams building a reconstruction technology: independent implementation in a real renderer, evaluation against the competing upscalers, and the SDK and bindings that get developers adopting it.

Neural reconstruction schematic: inputs, inference and image output.
Neural reconstruction schematic: inputs, inference and image output.

Your technology, in a renderer that is not yours

A reconstruction technology is judged inside somebody else's frame. The engine has to supply the inputs, manage the resources, place the work in the frame and deal with the output, and until that has been done against a real renderer the question of whether your technique is good is still open.

We do that implementation independently of the team that built the technique, which is the part that tends to surface what the SDK assumes. We integrated Arm Neural Super Sampling into our Breda framework and into Evolve this way, and where the hardware has not shipped yet we can work pre-silicon - that is where our hardware practice started, with a pre-silicon ray-tracing benchmark suite for Imagination Technologies.

Measured against what you are compared with

Your technology will be reviewed alongside DLSS, FSR and XeSS whether or not you choose the comparison, so it is worth having it made carefully first. We maintain Rust bindings to all four vendor interfaces, which means the competing upscalers can be run through one renderer and measured the same way: input preparation and integration cost as well as inference, and image quality on content that resembles what ships rather than a demonstration still.

Evolve, our GPU and AI benchmark, adds upscaling quality and performance workloads in its 2026 programme, including XeSS and FSR. That is a repeatable environment for the comparison, and an independent one - which is what makes the result usable outside your own slides.

Getting developers to adopt it

Silicon ships; adoption is a separate problem. A developer meets your technology through an SDK, a set of bindings, a sample integration and whatever the documentation does not cover, and the friction there decides how much of your hardware advantage ever reaches a player.

That is the same engineering our team does on shipping console and mobile titles, pointed at your adoption path: the integration a developer copies, the bindings that keep it current, and the platform and frame-budget work that decides whether it survives a real build. Our bindings are public and maintained, which is the form that work usually takes.

CrateWhat it reaches
arm-neural-graphics-rsArm's Neural Graphics SDK, including Neural Super Sampling
xess-rsIntel XeSS
fidelityfx-rsAMD FidelityFX, including FSR
nvngx-rsNVIDIA NGX, the interface behind DLSS

Traverse authors and maintains the bindings; each vendor authors the technology behind them.

Start a conversation · The Arm integration account · Evolve · Pre-silicon work.

A few useful answers.

Can you work pre-silicon?

Yes. Our hardware work began with a pre-silicon ray-tracing benchmark suite for Imagination Technologies, and we have run rendering workloads against unreleased architectures since.

Are the comparisons independent?

That is the point of them: we implement the competing upscalers through the same renderer and measure them the same way, which is worth more to you than a number only we can reproduce.

Does Traverse author the upscalers it compares?

No. Each vendor authors its own technology; we build the bindings, the integration and the evaluation around them.

Is inference cost the only measure?

No. Input preparation, integration, image quality and the complete rendering workload also matter, and a technique that wins on inference time can still lose the frame.

What has this looked like in practice?

We integrated Arm Neural Super Sampling into our Breda framework and into Evolve, connecting the SDK to the renderer's inputs, resource flow and output, and we publish Rust bindings for four vendors' interfaces.

Traverse Research + Traverse Games

Let’s turn a good question into working software.

Start a conversation