(The video compression on bluesky appears to be quite aggressive, I recommend watching the videos on the project page (ubc-vision.github.io/stochasticsp...) for full quality)

This was a fun and collaborative project with Shakiba Kheradmand (joint first author), @grgkopanas.bsky.social , Dmitry Lagun, Kwang Moo Yi, Mark Matthews and @taiyasaki.bsky.social .

Project page, PDF and higher quality videos: ubc-vision.github.io/stochasticsp...

Finally, the noise of the estimator can be significantly reduced using temporal anti-aliasing (TAA). It's likely that a sophisticated neural denoiser could further improve results.

We also provide an efficient and unbiased reverse-mode gradient estimator for stochastic transparency, enabling sorting-free training or fine-tuning. Most of this is not specific to 3DGS, and could apply other rasterized semi-transparent primitives too.

Our estimator naturally supports "pop-free" rendering, and optionally enables accurate volumetric intermixing.

Eliminating sorting allows fast, hardware-friendly rasterization. Stochastic transparency exposes a direct quality/speed trade-off in the form of the number of samples per pixel.

(2/2)

🏆 Our invited speakers for February to April, 2025 includes:

Dorian Chan, @axelparis.bsky.social, ZHEN XU, ezgi ozyilkan, Zhaocheng Liu, @deliovicini.bsky.social, Qi Guo, @niladridutt.bsky.social, Akshat Dave, Ethan Tseng, Ziyang Chen.

👉 For more details:
complightlab.com/outreach