Hi.
I am doing some post-production denoise experiments, let’s discuss about its applications and if it’s useful for anyone.
Using modern denoise algorithms we’ll have some kind of modern blur effects.
What do you think of this approach and what can it be good for?
Original BMW27.blend rendered image (default 20 samples).
Denoise #1:
Denoise #2, similar settings but a little sharper than #1 (look at the grill):
Stronger denoise level
Denoise #3:
Denoise #4, similar settings but a little sharper than #3 (look at the grill):
Note that #2 and #4 took longer to denoise, since much more stuff were running under the hood.
Here is an original render with 200 samples:
It looks pretty good. But what technique/software are you using? What are the time it takes for denoising?
It would be good to try other scenes too, like Sponza scene at 1000 samples
Using neural tensors but the code is far from being final, #2 and #4 need about twice as much time comparing to #1 and #3.
Right now it takes about 2s to denoise #1* but I think I can cut that down by twice or 3x. Unfortunately it’s not my main project, I use a little spare time on this one.
Where can I find that sponza blend file?
Classroom scene, rendered with 300 samples and shrinked to 960x540.
Denoised from 960x540 using #2 settings like the BMW27 example above:
You are having a very good result with denoiser you are using, considering it is a post processing on final render images. I’m really interested. Do you think this could analyze multiple frames for better denoising in animation?.
Also, I do not know if you had seen this thread (this is mainly for animation):
Why using a scene with unclear license? Lets just use free scenes instead. I will run it on all blender benchmark scenes at a later time.
It does a good job on fine grained noise, but on scenes with low samples no good results to be expected.
Classroom scene, rendered with 30 samples and shrinked to 960x540.
Denoised from 960x540 using #2 settings like the BMW27 example above:
Do you think this could analyze multiple frames for better denoising in animation?.
It works independently on single image/frame.
@YAFU
I took your image at http://www.image-share.com/upload/3039/198.jpg
Original
Denoised
As expected, I cannot denoise this well, since the noise is not fine enough.
>>Denoised with #2 settings. Wall-E image
Yes the denoising is good here! Not darker…Fine.
My conclusion: This is really just for very fine noise on any image.
Original fishy cat @ 1000 samples:
Denoised:
Ok, anyway seeing the others examples this looks a good denoiser you are using. I would like to know more about it. Not really technical details, but for example know about how difficult is to use it and abailable options for denoising.
for many years i have been using Gmic ( was GREYCstoration )
a pde for heat flow works great
– original
gmic 198.jpg -pde_flow 50,3 -sharpen 5 -o 123.png
and the result
side by side
the denoised is on the left
@YAFU:
At the moment I still use a pretrained neural network from someone else, so there will be a license problem if I publish it.
I need time to train my own network, complete the code and figure out what I can do with it, maybe as a Blender addon or something.
The usage is simple, just press “denoise” (if there will be such a button :)), no settings needed for the current state, since the best options seem to be #2 for all images.
Using neural tensors, the possibilities are endless, there can be much more improvements if I have time for this.
Original victor scene @ 600 samples:
Denoised:
