Cycles BCD - Bayesian Collaborative Denoiser POC

Hey, I spotted a proof of concept patch that allows to test a new denoise system that could possibly end up in Blender with Cycles.

Patch details: https://developer.blender.org/D3562

This is the denoise system used with Luxcore, But the guy (Élie Michel) wants more people to test the quality against Blender internal denoise system so Brecht can decide whether this would be worth adding as a second denoise system or even replace the existing system.

It’s a bit clunky to use right now as the BCD lib is not directly added to Blender yet, but to help out Ive done a windows build for people to test.

Instructions are here: https://github.com/eliemichel/bcd

Link to the BCD windows pre built tools are here: https://github.com/eliemichel/bcd/archive/v1.1.zip

Here’s the latest master from today with the Patch added: https://mega.nz/#!I8pgSagb!Zh4rNDayGOIuvH5sJe1OI5g2NZwCmxa3EIFGCS4mM_Q

2 Likes

I tried to test it yesterday. I was able to achieve everything related to Blender, but I could not build BCD-cli in Linux:
http://pasteall.org/1044435

I would suggest leaving a message on the guys git page, I don’t use Linux so cant help im afraid my friend.

hm so much i like the Denoiser, that we have now,sometimes it denoises to much details away.i wish there would be a better fine control.
if i look at your testrendering pics,the first things i noticed was,that the BCD denoiser,denoise even more details away.have you more testrenderings with different strength used?
details preserving is the key here,if it is not better than the Blender denoiser,i see not a great use for this.
just to be honest.

Disclaimer; Going by the current examples here, don’t have the build right now.

The BCD algorithm does a much better job in denoising regions completely missed by the current algorithm, but small details appear to suffer (unless we can somehow jack up the detail preservation at the potential cost of a less smooth result).

I think what Brecht hinted at in the patch thread could be the way to go, find a way to take concepts from both algorithms to create a new composite solution with far less issues and pitfalls.

1 Like