Machine-Learned Guide Channels for 3D Scene Denoising

The patent introduces guide channels as intermediary data structures that capture scene geometry and lighting information. These guide channels serve as mediators between the noisy path-traced image and the denoised output, enabling noise reduction without requiring additional light transport samples. The guide channels contain pre-computed information about surfaces, lights, and shadows that guide the denoising process.

The patent replaces the mechanical approach of increasing sample counts with a computational model-based approach using guided filtering. Instead of mechanically adding more samples to reduce noise, the system uses a learned denoising model that operates on guide channels and the noisy image, substituting brute-force sampling with intelligent signal processing.

The patent segments the denoising problem into multiple independent guide channels (e.g., depth, surface normals, shadow maps, lighting information) that can be processed separately. Each guide channel captures specific aspects of the scene, and the denoising model processes these segmented channels independently before combining them to produce the final denoised image.

The patent performs preliminary computation of guide channels during the rendering process itself, before denoising is applied. These guide channels are pre-computed with scene geometry and lighting information, so that when denoising occurs, the heavy lifting of scene analysis has already been done, reducing the computational overhead of the actual denoising operation.