Denoising Volumetric ReSTIR

Kaan Altıntaş
Code
Comparison of Volumetric ReSTIR: Base, OIDN, OptiX, Reference

Comparison of different denoisers on Complex Scene (an environment populated with fog and over 20,000 emissive triangles). From left to right: No Denoiser (BASE), OIDN, OptiX, and Reference. The annotated values indicate the frame time in milliseconds.

Abstract

This project investigates the trade-off between denoising strength and additional Volumetric ReSTIR sampling under a fixed compute budget. Real-time volumetric rendering pipelines often rely on denoising to suppress noise, but Volumetric ReSTIR itself is designed to produce low variance samples by reusing light transport paths across space and time.

The core research question addressed is: "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 Volumetric ReSTIR codebase with multiple denoising configurations (OIDN, OptiX) and performed controlled comparisons across different scenes and camera motions.

Video

Key Findings

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

Implementation

The work builds upon the original implementation of Volumetric ReSTIR by Lin et al. (2021). We integrated the following denoisers as render passes within the Falcor pipeline:

  • Intel Open Image Denoise (OIDN) – GPU
  • Intel Open Image Denoise (OIDN) – CPU
  • NVIDIA OptiX AI Denoiser (Updated to OptiX 9.0 SDK)

Scene Configurations

Simple Scene (Plume): A single volumetric plume illuminated by an environment map. This setup isolates denoiser behavior in participating media without surface occlusions.
Complex Scene (Amazon Bistro): A geometrically dense environment populated with fog and over 20,000 emissive triangles, introducing complex visibility, occlusion, and lighting interactions.

OIDN Settings
OIDN GPU Parameter Selection
OptiX Settings
OptiX Denoiser Parameter Selection

References

[Lin 2021] Lin, D., Wyman, C. and Yuksel, C., 2021. Fast volume rendering with spatiotemporal reservoir resampling. ACM Transactions on Graphics (TOG), 40(6), pp.1-18.

[Bitterli 2020] Bitterli, B., Wyman, C., Pharr, M., Shirley, P., Lefohn, A. and Jarosz, W., 2020. Spatiotemporal reservoir resampling for real-time ray tracing with dynamic direct lighting. ACM Transactions on Graphics (TOG), 39(4), pp.1-17.

[Andersson 2020] Andersson, P., Nilsson, J., Akenine-Möller, T., Oskarsson, M., Åström, K. and Fairchild, M.D., 2020. FLIP: A Difference Evaluator for Alternating Images. Proc. ACM Comput. Graph. Interact. Tech., 3(2), pp.1-23.

[Schied 2017] Schied, C., Kaplanyan, A., Wyman, C., Patney, A., Chaitanya, C.R.A., Burgess, J., Liu, S., Dachsbacher, C., Lefohn, A. and Salvi, M., 2017. Spatiotemporal variance-guided filtering: real-time reconstruction for path-traced global illumination. Proceedings of High Performance Graphics (HPG).

[Zhu 2023] Zhu, S., Zhang, X., Röthlin, G., Papas, M. and Meyer, M., 2023. Denoising Production Volumetric Rendering. ACM SIGGRAPH 2023 Talks.

[Kutz 2017] Kutz, P., Habel, R., Li, Y.K. and Novák, J., 2017. Spectral and decomposition tracking for rendering heterogeneous volumes. ACM Transactions on Graphics (TOG), 36(4), pp.1-16.

[Novák 2014] Novák, J., Selle, A. and Jarosz, W., 2014. Residual ratio tracking for estimating attenuation in participating media. ACM Transactions on Graphics (TOG), 33(6), pp.1-11.

[NVIDIA 2025] NVIDIA, 2025. NVIDIA OptiX AI-Accelerated Denoiser. Version 9.0.

[Intel 2025] Intel, 2025. Intel Open Image Denoise. Version 2.3.3.

[Falcor 2020] Benty, N., Yao, K.H., Clarberg, P., Chen, L., Kallweit, S., Foley, T., Oakes, M., Lavelle, C. and Wyman, C., 2020. The Falcor Rendering Framework.

[NRD 2025] NVIDIA, 2025. NVIDIA Real-Time Denoisers (NRD).