New Patent Removes Noise From AI-Rendered Images Faster Without Sacrificing Detail

By Patentlyze Team · Updated Jul 21, 2026

Ray tracing makes game graphics look gorgeous, but it produces noisy, grainy images that need AI cleanup. Nvidia has filed a patent for a smarter cleanup pipeline that sidesteps a problem the current approach doesn't even know it has.

What Nvidia's denoising workaround actually does

Imagine a photographer who shoots in low light and uses software to sharpen the grainy result. Now imagine that before the sharpening runs, someone smears a filter across the photo. The sharpener does its best, but it keeps fighting the filter instead of the grain. That's roughly what happens today when AI cleans up ray-traced game images.

Ray tracing (the technique that makes light, shadows, and reflections look photorealistic) produces images full of random speckles. An AI denoiser smooths those out. But many visual effects, like lens flares, depth-of-field blur, or bloom, are painted on top of the image before the AI sees it, which confuses the cleanup process.

Nvidia's patent describes a way to first strip those effects off, hand the cleaner image to the AI denoiser, and then record exactly what was removed so it can be reapplied afterward. The result should be sharper, cleaner final frames with fewer artifacts.

How the effect-subtraction pipeline feeds the neural denoiser

The patent describes a three-step pipeline inserted between a path tracer and its AI denoiser.

• Step 1 (original image): A path-traced frame is captured with all its raw noise intact.
• Step 2 (effect applied): A screen-space effect (meaning a visual effect calculated in 2D pixel space rather than in the 3D scene itself, like bloom, motion blur, or chromatic aberration) is layered on top.
• Step 3 (difference image): The system computes the mathematical difference between the original noisy frame and the effect-layered version. This difference image tells the denoiser exactly what was added.

Both the effect-layered image and the difference image are then fed together into a neural network denoiser (an AI model trained to distinguish meaningful image detail from rendering noise). Because the denoiser now knows what the effect contributed, it can clean the underlying image without mistaking the effect for noise or trying to smooth it away incorrectly.

The key insight is that screen-space effects applied before denoising act like a second source of confusion on top of the existing noise. By making that confusion explicit and measurable, the AI can handle both at once.

What this means for real-time ray-traced graphics

Ray tracing at real-time frame rates still requires heavy AI assistance, and any artifacts the denoiser introduces are immediately visible to players. If Nvidia's approach works as described, it could reduce the telltale smearing or ghosting that occasionally appears around light sources and reflective surfaces in current ray-traced games.

This is also consistent with Nvidia's broader strategy around DLSS (its AI upscaling and denoising suite). Better denoising quality means developers can use fewer path-traced samples per frame, which is the main knob that controls how expensive ray tracing is to run. In other words, cleaner AI cleanup can translate directly into higher frame rates for you at the same quality level.

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