← Ruihan Yu

ACM Transactions on Graphics  ·  2026

Best papers award honorable mention

Sample Matching for Joint Extinction Gradient
Estimation in Differentiable Volume Rendering

Ruihan Yu1,2 · Yu-Chen Wang3 · Jingwang Ling4 · Feng Xu2 · Shuang Zhao4

1The University of Tokyo    2Tsinghua University    3University of California, Irvine    4University of Illinois Urbana-Champaign

Paper teaser: DRT vs Reference vs Ours, with error heatmap insets
A single shared probe per segment (Ours) reduces gradient variance by up to 80% over DRT on the same compute budget.

Live, in-browser demo of our method below — scroll on and try it

TL;DR

We model the extinction (or, plainly, density) gradient as the difference between the in-scattering radiances of two paths. That mirrors what the physics is doing: increasing density at a point adds in-scattering and attenuates background transmission, so the gradient is naturally the sum of two terms with opposite signs.

Existing unbiased estimators never couple these two terms — they evaluate the scattering and transmittance contributions at different points along the segment, so the cancellation that should reduce variance becomes pure noise instead.

Coupling them is non-trivial because the two terms live on different integration domains. A short derivation built on Fubini's theorem and the definition of transmittance is enough to put both terms on a shared domain, so they can be evaluated at the same probe. Once they share a sample location, their opposite-sign correlation collapses much of the variance.

In practice this is a near drop-in: even without the most-tuned codepath, a handful of lines on top of an existing volumetric path-integral estimator already yields a large variance reduction.

Why two opposite-sign terms?

Nudging the medium denser at one point does two things at once: it blocks more of the red, in-line background radiance the camera was already seeing, and it scatters more of the blue, off-axis radiance into the same outgoing ray. The density gradient is the difference between the two radiances. Drag \(\sigma_t\) below to watch the trade-off.

Where does the scattering probe land?

The three estimators agree on the transmittance probe (uniform on the host segment) and only differ in where the scattering term reads \(\partial\sigma_t\). Each schematic shows two lines on the same axis: the top one is the scattering probe with its sampling PDF drawn as a blue silhouette; the bottom one is the transmittance probe. DRT's scattering PDF lives on the segment extended to the medium boundary \(x_j^{\bot}\) (violet dashed).

Per-cell variance, \(N{=}128\) grid, 128 spp
scene
density \(\sigma_t\)    albedo
FF
DRT
SM
low variance high variance

Video

BibTeX

@article{yu2026samplematching,
  title   = {Sample Matching for Joint Extinction Gradient Estimation in Differentiable Volume Rendering},
  author  = {Yu, Ruihan and Wang, Yu-Chen and Ling, Jingwang and Xu, Feng and Zhao, Shuang},
  journal = {ACM Transactions on Graphics},
  year    = {2026}
}