← All ideas
  • Published 21 Dec 2023

A better denoiser

arxiv.org/pdf/2310.03507.pdf

Well, this research paper seems to have gone under the radar. It even used Cycles for testing.

100%

Development Status: Open

Upvote 7 Downvote 0
fungpomin
21 Dec 2023

4 Comments

would be really useful! I've gone to render things in the past and found that at specific lighting angles and with specific roughness values in the texture, the denoising simply didn't know what to do with it, and created a mushy, noisy result

that's caused by adaptive sampling creating patterns in areas of sparse rays, then the denoiser rightly considers those patterns intentional. Turning off adaptive sampling fixes that.

zuorion

In case link dont work (like for me), this leads to:

RL-based Stateful Neural Adaptive Sampling and Denoising for Real-Time Path Tracing Antoine Scardigli, Lukas Cavigelli, Lorenz K. Müller

synthical.com/article/aeeaeef9-8cfd-4396-8d06-e3e198bd3269

I develop an addon for Cycles called Turbo Tools. It generally give a 3 to 5x render time reduction, compared to the 1.6x in that paper. It comes with a bunch of additional tools to speed up the entire rendering and post production tasks too 👍

blendermarket.com/products/turbo-tools-v3---turbo-render-turbo-comp-temporal-stabilizer

It would still be great to see the sampling improvements in that paper added to cycles though.