Please do not use this thread for general feature requests or requests for help. It is intended to support Cycles X development. See the site header text for links to platforms for help and requests.
Please try to refrain from comparisons with other renderers. This thread should be focused on Cycles X performance, development and possible issues.
Please only tag a developer username in your post if you’ve got a relevant Cycles-related issue or question that needs to be answered by a specific developer. The Cycles developers are reading this thread and responding regularly, and frequently being tagged in posts is bothersome.
@PetrTarasov@AndreasResch
NLM is not removed, it is simply not added at moment.
In the presentation paper it is up for discussion to add it again or reject it.
What lsscpp said, you want to denoise all the passes individually.
Some of our production shots we’re also doing denoises in other color spaces to force details to be retained. In comp we also often denoise and do transformations in Log-space to avoid/reduce filtering artifacts as some of the filter types can be quite harsh on float images.
The good thing about this approach is that you can even do things like subtly sharpen the albedo to squeeze even more “perceived” sharpness out of the renders.
In addition to spatial denoising we’re also emplying temporal denoising by using the vector pass to interpolate between 3 or 5 frames and conducting a median. That way we get very little noise and very stable denoising.
To do this job with OIDN user must have enough RAM on high res(4K,8K) image via compositor. Anyway NLM is not perfect but retain details better, have more options to control result.
Could we maybe see that in actual Blender nodes? Or at least label more clearly the mix types and such? I’ve been using a node group I found online for quite some time that I suspect is not ideal.
Subsurface scattering seems broken in Cycles-X, at least with denoising done in the compositor:
SSS at 0.03
SSS set to 0.0
If you have output set to RGBA instead of RGB, that black crap all ends up transparent right through to the checkerboard pattern (ignoring all objects in the scene).
It’s not the best case and not a real scenario but it clearly shows improvement over older OIDN, also shows that NLM is still sharper (but ugly on edges). Also OIDN was slowest from all methods.
Maybe there’s better and newer methods to denoise using the same techniques like NLM?
Hi, Cycles-X use a different Optix denoiser, I cant remember where I read it but Nvidia provide more details with Optix 7.3.
Can you compare your file with it, please?
I use all denoisers, depends on the scene.
And a bonus tip that many don’t know: placing a dot in front of the nodegroup’s name makes it invisible. Useful to not overcrowd the nodegroup category when you add them via the menu
I’ve just downloaded cycles X the other day and its amazingly fast. The main problem I have at the moment in that of storage. I’ve noticed blender has gone up to 5 or 6 gb and as a person with not very good storage its not the best. Ionly noticed this though when I was unusually low in storage which flagged the mac “your system has run out of storage” window so this might be normal. Another thing is that final render doesn’t work and only just shows transparency. I’m pretty sure this is a bug so i’ve filed in a bug report. Also when ever I render it does not update any change in samples, I have to save and close to update. Overall I am really excited for cycles x as I use Cycles a lot as Eevee can get quite laggy at times.