Denoising Volumetric ReSTIR

The work builds based on the original implementation of Volumetric ReSTIR by Lin et al.(2021). While this technique already produces low variance samples by reusing light transport paths across space and time. In practice, however, real-time volumetric rendering pipelines often rely on denoising to further suppress noise.

This raises a key research question:

"Under a fixed frame-time budget, is it more effective to spend compute on stronger denoising, or to allocate that budget to improved Volumetric ReSTIR sampling?"

To answer this, we extended the original codebase with multiple denoising configurations and performed controlled comparisons across different scenes and camera motions.

• Perceptual Quality: Denoisers significantly reduce perceptual error, especially in complex scenes.
• Simple Scenes: In simple volumetric scenes, Volumetric ReSTIR already produces clean samples, making aggressive denoising less beneficial.
• Compute Allocation: Reallocating compute from denoising to sampling can recover much of the quality lost by weaker denoising.
• Efficiency: Low quality denoising combined with improved sampling often matches and sometimes exceeds the quality of high quality denoisers.
• Optimization: Selecting the strongest denoiser is not always optimal; best results come from balancing sample quality and denoiser strength.