TUM's HaloRT Renders Clouds 5000x Faster by Ditching Path Tracing
TU Munich researchers introduce HaloRT , inferring volumetric radiance without any path samples. Renders clouds in about 10 ms versus 49 second path-traced references on an RTX 3090. An attention-augmented U-Net consumes optical depth, transmittance and scattering features from primary rays only.
Takeaways
- TU Munich researchers introduce HaloRT , inferring volumetric radiance without any path samples.
- Renders clouds in about 10 ms versus 49 second path-traced references on an RTX 3090.
- An attention-augmented U-Net consumes optical depth, transmittance and scattering features from primary rays only.
- Beats MRPNN, OIDN and a specialized volumetric denoiser on LPIPS and temporal stability.
- Cost stays roughly constant as density or camera zoom changes, unlike path tracing.
- No code or weights released; PSNR trails denoisers due to small global intensity biases.
Volumetric path tracing consumes large sample budgets because light can scatter repeatedly inside clouds, smoke, and fog. Each stochastic event adds variance, so dense media may require thousands of samples per pixel before an image converges.
A new paper from Ludwic Leonard and Ruediger Westermann at the Technical University of Munich proposes a faster route. Their method, HaloRT , predicts integrated pixel radiance from inexpensive ray and lighting features. The authors report roughly 10 ms per image on an NVIDIA RTX 3090 for scenes whose path-traced references take 49 to 51 seconds to converge.
Why clouds drain sample budgets
Monte Carlo volume renderers trace rays through a participating medium, sample scattering locations from its density field, and average many paths per pixel. Dense clouds produce long, complex paths with high variance. Denoisers reduce visible noise after a low-sample render, though generating that input often remains the dominant cost.
Earlier neural methods make different compromises:
| Method | Approach | Constraint |
|---|---|---|
| MRPNN | Learns density-conditioned transport | Requires many environment samples and can remain noisy |
| Intel Open Image Denoise | Processes rendered color and auxiliary buffers | Can smooth fine volumetric structure |
| Specialized volumetric denoisers | Target noise in participating media | Retain variation from their stochastic radiance input |
Four feature groups carry the signal
HaloRT derives its runtime inputs from primary rays and inexpensive lighting evaluations. The feature set describes the volume encountered by each camera ray and the amount of illumination that can reach it:
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
