Ragged Neighborhood Attention Removes Render Noise While Preserving Layers

RaNAD uses ragged neighborhood attention to denoise deep Monte Carlo renderings across space and time while preserving layered compositing structure.

Paper 《Ragged Neighborhood Attention for Spatiotemporal Neural Denoising of Deep Monte Carlo Renderings》

Authors Xianyao Zhang · Gerhard Röthlin · Tunç Ozan Aydin · Farnood Salehi · Marios Papas

Research Disney Research Studios

Presented at SIGGRAPH 2026

Release 2026-07-16

Video 6 minutes 32 seconds

A Spatiotemporal Denoiser That Preserves Deep Image Structure

Deep images store multiple samples at different depths within a single pixel in the form of ‘bins.’ This makes it easy to separate overlapping objects for compositing, but because the number of bins varies from pixel to pixel, deep images do not work well with conventional neural denoisers that use fixed convolution kernels, making Monte Carlo rendering noise difficult to remove.

The research team proposes Ragged Neighborhood Attention, or the RaNA operator, which dynamically identifies the relationships between bins needed when examining spatial and temporal neighbors. Based on this operator, they built ‘RaNAD,’ the first spatiotemporal neural denoiser for deep Monte Carlo renderings.

RaNAD uses multiple U-Nets, a multiscale reconstruction architecture, and an optimized CUDA implementation. It supports both deep-Z and deep-OID formats and processes up to seven frames at once, reducing temporal flicker and instability compared with single-frame methods.

Even after denoising, the resulting images preserve the layer structure needed for deep compositing. When flattened into planar images for comparison, they also achieve quality competitive with powerful general-purpose image denoisers, pointing toward a workflow that can connect directly to offline rendering and compositing pipelines using multiple AOVs.

🏷 Other posts for the same work →

Maru