DeepDenoiser for Cycles

(DDB) #121

this could be amazing if this denoiser is implemented into 2.8? anyone know ? i didnt read all the post
the devlopper who did this is amazing



Thanks for the compliment :blush:

A lot of work is needed before it could even be used in Blender. Right now, I still have to focusing on improving it. It still takes quite some time.



Here are two more denoising examples of well known images:


(BlackRainbow) #124

ok… now this is getting serious


(Romanji) #125

Yep. It already looks like it surpasses the quality you get from Nvidia’s Optix denoiser, or at least it is on par with it.


(zeealpal) #126

Wow! This is looking impressive!

Which passes does it require to work, would they have to be from Cycles specifically?

1 Like


Thanks for the positive feedback!

@zeealpal, it denoises one pass after another and requires obviously that pass, but also the normals and some kind of “flags” to indicate what the pass is. That’s why it is very Cycles specific. Theoretically, there are ways to make it more general, such that it can be used for other renderers or even for other kinds of noise removal. However, that is not what I am planning to work on.


(DDB) #128

look amazing :heart_eyes:
is it used during the render like right now how could it be used also after in the compositor ?
if it is used during does it take vram memory ?



The denoising does not take place during the rendering, but afterwards. All passes are stored after the rendering and one after the other gets denoised. Right now, this is done manually, but I hope to be able to have a node in the compositor where you denoise all or just some selected passes to give artists as much flexibility as possible.


(doublebishop) #130


Do you think it would be possible to integrate it in like the current denoiser is? as in switching out the current version for this version and have it happen seamlessly in the background?


(DDB) #131

yes , a denoizer as a node is excellent



As far as I have seen, there are changes for the denoiser for 2.8. I will do my best to create a seamless integration for users (as every developer). I don’t want to say how it is going to look yet, because I don’t know and because it is not going to be my decision alone.


(zeealpal) #133

Thanks, will be great to use with Cycles anyhow!

My thoughts were to how applicable it may be to Luxcorenetnode, which is quite similar to Cycles, except its current denoiser isn’t as good as Cycles, or this :slightly_smiling_face:



As mentioned before, right now it is specifically made for Cycles, but it could be adapted for other renderers. It would require a decent amount of training data as a starting point. I am open to help other developers with it, but it is not on my TODO list.


(noki paike) #135


in the future a deep learning algorithm will recognize the grease pencil shapes and rebuild the 3d version of them

a demo shown by google has already shown this possible



That’s indeed a very cool project with an immense potential. Going from a 2d image/photo/scribble to a 3d model relatively easily would empower a lot of creativity. What I most like about it, is the obvious ways to integrate it into content creation workflows.
You could start with one drawn shape and a 3d model would be created. Now you inspect it and you find that the side view should be different. So you switch to the side view and let it automatically create grease pencil shapes, such that you can easily modify them as you want. Now, you can update the 3d model based on those two shapes to get a result that is closer to what you envisioned and you can continue as needed.
I am a huge fan of this sorts of interactive workflow where artists have the full control and yet can focus on the actual art. This would without doubt encourage artists to try more ideas out compared to what can currently be done.

As machine learning is heavily overhyped, I feel the need to mention very clearly here, that the discussed project is likely years away from being practically usable.


Idea for AI and Cycles Render!
(kakapo) #137

would it also be possible to skip cycles? :slight_smile:

to just train it with the scene data (geometry, materials, lights) and the final high quality render?

1 Like

(Piotr Adamowicz) #138

Probably not. AIs are too dumb for that. They sort of have this monkey-see-monkey-do mentality :wink:

It’d be like trying to teach a mouse to paint from description, without reference :stuck_out_tongue:


(noki paike) #139

I agree. immense potential.


(Photox) #140

This is absolutely not true. GANs have been generating photoreal images for years, and I personally wrote a blender script that generated cubes rotated at different angles and trained a conv net to generate images of them based only on the angles. It’s only a question of time and speed.