Optix AI Denoiser


For everyone interested, here is a video from @Bone-Studio demonstrating the OptiX denoiser plugin from Remington Graphics (don’t know whether Grant Wilk is in this forum…).

(Riley Brown) #49

Just an update for everyone here. Latest news is Grant getting ready to release the plugin Friday!

It’s been cool to see the community in action developing this plugin. Here are some results from the final version on the benchmark. 4 samples.

(English is not my native language) #50

Linux plugin?, binaries or how to compile/build?

(Riley Brown) #51

Sorry, code is not my native language. :wink:

Hit @ RemiGraphics up on Twitter I’m sure he’d be happy to answer for you.

(English is not my native language) #52

OK thanks
I do not have a Twitter account… I do not have tattoos either. What kind of uncommon person I am in these current times :smiley:

(Shylon) #53

(post withdrawn by author, will be automatically deleted in 24 hours unless flagged)

(yandrychoy) #54

Hi, impressive denoiser, it will be free? Thanks

(captainkirk) #55

Yes, it’s going to be a free addon.

(Thesonofhendrix) #56

This will hold us over until we get Deep Blender.

(DDB) #57

is this running in real time?

(Riley Brown) #58

It’s out tonight, and it’s free thanks to Remington Graphic’s and their hard work. Go give it a try!

(rboxman) #59

Really cool, seems to work here on some simple scenes I’ve thrown together.

However, I may have found some cases where the denoiser produces problematic results at higher sample counts, worse than the initial render (ironically). Where/how can we file issues for these – probably requires more training?

(Riley Brown) #60

Github link for reporting issues/bugs: https://github.com/grantwilk/DNOISE

(rboxman) #61

Cool. Looks like someone else beat me to the punch though by a few minutes by filing effectively the same issue I believe.

(moony) #62

I tried it on the BMW benchmark scene with 70 samples - and to be honest, I think the inbuilt Cycles denoiser does a much better job at controlling noise and preserving fine detail (Optix is the top one - click to expand the image - then use the cursor keys to fick back and forward).

edti - hmm, looks like it’s quite dependent on the options you check. For the above test I checked both the HDRI and extra passes checkbox. If you don’t use these, the render is much better:

(Ace Dragon) #63

I’m surprised the thin highlights look so cruddy with the AI denoiser, and here I thought the Cycles denoiser produces a few artifacts there.

The bottom image is better in the lights, but the windshield is almost opaque and part of the mirror of the second car was not denoised at all.

Is it possible that Nvidia again used marketing magic here? At least there’s DeepBlender’s denoising work (in which the actual results in difficult scenes give the appearance of actual magic).

(Peetie) #64

Is it me or is the Blenders own Denoiser significantly better? I tried 12 different scenes and the difference ( not denoised vs denoised) is minimal. For those 2 a 3 seconds I render a bit longer with my GTX 1080 TI.
Or could it be I am doing something wrong?
Tried around 100 samples and 20. HDRI training on/off and extra passes on/off (which are giving artifacts all the time).

And when using pictures is not denoising, but only makes the darker pixels brighter.


My tests showed that they have different strengths.
Blender’s denoiser is really good at retaining details while having more trouble with large smooth areas where it produces very cloudy results.
OptiX on the other hand has a tendency to overblur the details away while looking amazing on large smooth areas.

Edit: I only used the default settings, even though the results could definitely be improved with other settings.

(Peetie) #66

Sorry to say but I believe but the denoise node from bwide node or even despeckle node does a better job. Really.
Could it be the marketing?


Feel free to show the comparisons to backup your claims. I would be surprised if you could achieve better results against actual denoisers without a massive time investment.

This image is pretty representative for the tests I have done.