Nvidia Patents an AI System That Scrubs Grain from Ray-Traced Images

By Patentlyze Team · Updated Jul 21, 2026

Ray tracing makes 3D scenes look stunning, but it also makes images grainy and noisy unless your GPU does enormous amounts of extra work. Nvidia's new patent describes an AI that scrubs that noise away, faster and more accurately than existing methods, by letting the network peek at neighboring parts of the image while it works.

What Nvidia's ray-tracing cleanup actually does

Imagine your GPU is drawing a realistic 3D scene by simulating actual light bouncing around a room. That process is slow, so the GPU cuts corners and fires fewer light rays than it really needs. The result looks great in some spots but grainy or blotchy in others, like a photo taken in a dark room.

Nvidia's patent describes a system that steps in after those initial calculations and uses an AI to "fill in" the missing detail. The AI looks at small tiles of the image and, crucially, lets each tile peek at the tiles around it before deciding how to clean itself up. That cross-tile awareness is the key idea here.

The underlying AI model is a transformer network, the same family of models that powers large language models like ChatGPT, but repurposed for images. By giving it overlapping "attention windows" that extend beyond each tile's edges, the system avoids the telltale seams and blotchy patches that simpler denoising tools can produce.

How the attention windows reach across image patches

The patent covers a denoising pipeline built on transformer networks, a type of AI architecture that weighs relationships between many pieces of data at once (originally designed for text, now widely used for images and video).

The process works in steps:

• The GPU runs path tracing (a physically accurate but noisy light-simulation method) to produce a rough, grainy frame.
• That frame is divided into small rectangular patches, which the AI processes in parallel.
• For each patch, the system defines a windowed attention region that extends beyond the patch's own borders, so the AI can reference pixels from adjacent patches while cleaning up each one.
• The AI then produces an updated image where noise is reduced while edges and fine detail are preserved.

The critical technical claim is that windowed attention regions exceed the patch boundaries. Standard tile-based denoising keeps each tile isolated, which can leave visible seams. By letting the attention window bleed into neighboring tiles, the network has more context and can make better guesses about what the true, noise-free pixel values should be.

The patent's claim covers this approach at the processor or circuit level, meaning it is describing hardware-accelerated execution, not just a software algorithm running on a general CPU.

What this means for real-time ray tracing in games

Ray tracing is already standard in high-end PC gaming, but it is still computationally expensive enough that developers have to lean heavily on denoising to make real-time frame rates possible. Nvidia's existing DLSS system already does some of this work; a patent like this suggests Nvidia is continuing to invest in the AI side of the denoising pipeline rather than just throwing more GPU cores at the problem.

For you as a gamer or 3D artist, better denoising means cleaner images at lower rendering costs, which could translate to higher frame rates, lower power draw, or more realistic lighting in scenes that would otherwise be too expensive to render. It also matters for film and product visualization studios that use path tracing for offline rendering, where faster denoising cuts production time.

Editorial commentary on a publicly published patent application. Not legal advice.