Computer Graphics

Denoising

Trading variance for bias so that low sample counts are shippable.

Colour is the family; a dashed line is the second member of it.

Drag to pan · scroll to zoom · click a node to open it
G a-trous-filter A-Trous Wavelet Filter bilateral-filter Bilateral Filter a-trous-filter->bilateral-filter wide support in a few sparse passes, not one big kernel monte-carlo-integration Monte Carlo Integration bilateral-filter->monte-carlo-integration trades a little bias for a large drop in variance firefly-clamping Firefly Clamping firefly-clamping->monte-carlo-integration rare huge samples never average away in practice neural-denoising Learned Denoising svgf Spatiotemporal Variance-Guided Filtering neural-denoising->svgf learns the filter from data instead of hand-tuned weights backpropagation Backpropagation neural-denoising->backpropagation the filter is learned from noisy and converged pairs svgf->a-trous-filter drives filter width from measured variance temporal-reprojection Temporal Reprojection svgf->temporal-reprojection the variance estimate comes from accumulated history

5 nodes

A-Trous Wavelet Filter

Applies a small kernel repeatedly with exponentially increasing sample spacing, reaching a wide filter in a few passes instead of one enormous kernel. · 2010

Learned Denoising

Trains a network on noisy and converged pairs, taking auxiliary buffers such as albedo and normal as input. Reconstructs detail that hand-tuned filte… · 2017

Spatiotemporal Variance-Guided Filtering

Estimates per-pixel variance from the temporal history and drives the spatial filter width from it, so converged regions stay sharp while noisy ones… · 2017

Bilateral Filter

Weights a blur by similarity in depth, normal and colour as well as distance, so it smooths within a surface but not across its boundaries.

Firefly Clamping

Caps individual sample contributions at a threshold. Biased, unapologetically so: a single sample thousands of times brighter than its neighbours wil…