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

Disney Research Studios has unveiled ragged neighborhood attention, or RaNA, a new operator built for semi-structured deep images where each pixel can contain a variable number of bins. That irregular layout has been a bad fit for fixed convolution kernels, which is one reason deep-image denoising has lagged behind flat-image workflows.

RaNA dynamically resolves bin-to-bin relationships across both spatial and temporal neighborhoods, letting the team build RaNAD, a spatiotemporal neural denoiser for deep Monte Carlo renderings. The system uses a multi-U-Net backbone, multi-scale reconstruction, and an optimized CUDA implementation, and it supports both deep-Z and deep-OID data.

The practical detail for production teams is that RaNAD keeps the layered structure needed for deep compositing instead of flattening everything into a single image too early. It also handles temporal windows of up to 7 frames, which should help stability in animated sequences and reduce flicker in offline rendering pipelines.

For developers working on rendering, compositing, or ML-assisted image cleanup, the interesting part is the kernel-based approach: it can denoise multi-AOV images efficiently while still improving quality substantially over prior deep-image denoisers. When flattened for evaluation, it also lands in the range of strong flat-image denoisers, which makes it more than just a niche deep-compositing tool.