Innobright Altus Denoiser Comparison shots

[FrostKiwi]

Hey ho.
Some Comparison shots and thoughts about Innobright's Altus Denoiser.
It's a denoising tool, recently released with corona support. It promises shorter render times, by completly elimintaing noise with a smart algorithm designed for monte carlo sample renderers.
Original thread by innobright here: https://forum.corona-renderer.com/index.php/topic,10576.msg

disclaimer: my humble opinion

tl;dr
good previz at 10+10 or even 5+5 passes. But for a good end result you need 50+50 passes, at that point you might just aswell render one 100 pass image and couple extra for the time it would take you to setup and actually filter. Corona Adaptivity to the rescue, hopefully soon.
+Magically Clears reveals a near finished image even with low passes
+GPU support, with 34secs of filter time not bad
+Way better than NeatVideo approach, some people have, as it 100% keeps edges untouched
-Texture detail blurring unavoidable, blurs shadow lines
-Requires workflow change and demands two frames to be rendered, Animation included, meaning two identical animations with differnet noise
-Typical Filtering splootches, unusable at lower passes with animation
-One more thing to buy


Summary:
It is not the magic algorithm I saw in 2011 in a paper by AGL. It's still pretty awesome for previz stills, but the workflow does change and in Animation I do see some jumping splootches produced by filtering. It DOES lose fine Texture detail, like small speculars in the wood texture. At lower passes it's unusable due to blurring the shadow lines and general sploochiness. Here the Video from 2011 by AGL and their magic filter algorithm, and Comparison shots below. IT IS the best I have ever seen for filtering that actually released, but it is still no good enough for me. At high resolution and especially for previz, this may be good. Also there is the rendering speed penelty for including all the passes, which doesnt matter for final render, since they are needed, but for previz it is noticible.

Workflow based on official Altus Corona Documentation pdf
All Filtering for 1280x720 images were performed on a fx-8350 4.5GHz + gtx 770 and always took 34 seconds.
All Default rendering settings, exept image filtering, which was set from tent to none. UHD cache was used.
Usually I do settings differently for faster sampling, like more rays, less samples, lower MSI, but in this test case, all default. Rendering 100 passes is madness if no dof or moblur is being used.

Images go as follows: Unfiltered vs filtered vs unfiltered with the next higher amount of render time

[johan belmans]

Thanks for sharing your opinion/research.
My choise will be: wait for Corona's denoise or less noise solution.

[Juraj]

Nice tests !

I personally do find the filtered100 quite good looking.

I find the need for filtering to occur at high resolutions and for complicated scenes (I can have noise in certain scenes far above 1000 passes), two factors that can multiply rendering time into eternity (30 hours on dual-xeon rig).
I believe the "smudiginess" wouldn't be so bothersome at such resolution as I use heavy filtering anyway (3+ px filter).

Maybe you could try at least with 3840px ?

[FrostKiwi]

3840px ?

Nah, I don't have a freakin' dual xeon rig :P

This is indeed true for higher resolutions. Why the 3px Filter though? Isn't that overkill?
Also here is 200passes filtered. I can't accept the blurring of the wood on the floor (especially vs twice as long rendertime below)
You may have more room to play around, but I do animation and rarely render at high resolutions. That blurring flickers btw. very Slightly, but still does.

[romullus]

Those are really nice tests!

For those who are obsessed with noise, this Altus thingy might be Godsend, but i always like small amount of noise in my images. For my highly untrained eye, noise equals details and big flat completely noise free areas looks very artificial. And still, i wait for Corona's adaptivity with big anticipation :]

[FrostKiwi]

For those who are obsessed with noise, this Altus thingy might be Godsend, but i always like small amount of noise in my images. For my highly untrained eye, noise equals details and big flat completely noise free areas looks very artificial. And still, i wait for Corona's adaptivity with big anticipation :]

Same - back in the day, I kid you not, I did AO passes with lower amounts of samples to let the darker areas have noise, before I did noise adding. Worked wonders for vfx.

[Chakib]

Those are really nice tests!

For those who are obsessed with noise, this Altus thingy might be Godsend, but i always like small amount of noise in my images. For my highly untrained eye, noise equals details and big flat completely noise free areas looks very artificial. And still, i wait for Corona's adaptivity with big anticipation :]

Amen !

[Juraj]

3840px ?

Nah, I don't have a freakin' dual xeon rig :P

This is indeed true for higher resolutions. Why the 3px Filter though? Isn't that overkill?
Also here is 200passes filtered. I can't accept the blurring of the wood on the floor (especially vs twice as long rendertime below)
You may have more room to play around, but I do animation and rarely render at high resolutions. That blurring flickers btw. very Slightly, but still does.

Indeed, the clean rendered one is far superior than the filtered, I just wondered how much of that would still be issue in high-res, where less of texture filtering happens since the scene/display ratios get more equal.

Also, is there no configuration of any sorts ? Aggresivity of the filter ? Or percentage of filtering, does it need to do a 100perc. job ? Seems the current solution is too much of one single box for all type of shoes...

Filtering is not godsend only for noise obsessed but possible evolution of raytracing imho. Adaptivity is nice, because it will make my 36 hours render into 20 hours. OK, but that's not it ;- )
Also, there is very little threshold for CGI noise in high-end visualization, where clients use 100mp Medium format cameras in studio setups, producing incredibly clean, high-res photography. They want the same look. And I agree, CGI noise doesn't look anywhere as nice as real filmic noise. Even the distribution plays strongly against it, pushing it into uncanny valley.


(ps: the 3px filter, not at all, at 8k, it's just right. I use default 1.5 rarely, mostly for sharp 2560px renders, do moment I go higher, it's 2px. And 3px for very high-res to avoid CGI look, otherwise you would see too much of an texture detail, which isn't so nice unless you use extremely texel-dense maps. Also geometry detail rarely holds up to such resolution, and the blurring makes it smoother, nicer. The image is still very crisp, important for large book prints)

[denisgo22]

It is all very well,  but try to remove
noises after DOF or Motion Blur in Altus, that especially takes the most time of rendering,
and your will see that the result is poor and not correct///

[Noah45]

All stills with Big C, so far, but the roaming patchiness in animation is not encouraging for Altus. Come on Adaptive Corona Render

SairesArt, thx for 1st and bestest review

[FrostKiwi]

Also, is there no configuration of any sorts ? Aggresivity of the filter ? Or percentage of filtering, does it need to do a 100perc. job ? Seems the current solution is too much of one single box for all type of shoes...

All filter settings in Image.
From Documentation:

Code: [Select]

-r [ --radius ] arg (=10) Filter radius. default value is 10.

--kc_1 arg (=0.449999988) This value is sensitive to image detail, but less aggressive towards noise. Default value is 0.45.

--kc_2 arg (=0.449999988) This value is the same as kc_1, but for a second pass. Default value is 0.45.

--kc_3 arg (=1e+010) This value determines aggressiveness towards noise, but may cause loss of detail. Default value is 1e10f.As you see, these settings are not well documented and good luck guessing what they do without trial and error :|
There is a Corona "Preset" it does not change these values, but I'd love to know what it does.

[maru]

Thanks for the tests, it must have costed you some time!

It looks like the blurriness of floor texture and carpet is not acceptable. I wonder if it could be somehow optimized - maybe different Corona settings would affect it (filtering, filter size?).

I think it is more important to understand how to set up Corona properly. For example in the images from the 1st post it looks like GIvsAA is wrong. There is still a lot of noise even after 200 passes. 200 passes is probably too much, and it should be fairly easy to determine how much AA is required - for example if fine details look ok after 50-100 passes, then this is the goal number of passes, and GIvsAA should be adjusted accordingly (increased) to produce clear GI.

Corona's bucket engine has some basic adaptivity, so it could also speed up rendering, and I bet it would produce less artifacts than the denoiser.

Of course there is little chance that:
a) I am wrong
b) we don't know how to use the denoiser properly due to its poor documentation

[FrostKiwi]

I think it is more important to understand how to set up Corona properly. For example in the images from the 1st post it looks like GIvsAA is wrong

-->

Usually I do settings differently for faster sampling, like more rays, less samples, lower MSI, but in this test case, all default. Rendering 100 passes is madness if no dof or moblur is being used.

yeah, I know :P
I think comparisons should be setup with default settings, especially since Corona's principle is simplicity and "default settings are good enough in 80% of scenes"

Also, the benchmark scene had that odd GI shadow on the left of the image on the right side if the door frame cast by the black Bookshelf, that wouldn't Cleanup, only after a long time. Neither in the benchmark back in the day, nor now. Maybe compatibility issue with old scene or MIS failure.

Since I do animation I found the sweet spot to be 8 passes. (Simple geometry) (Without dof or moblur) and for final render just crank GI vs AA to 64 or higher. With simple geometry 8 is plenty.

It looks like the blurriness of floor texture and carpet is not acceptable. I wonder if it could be somehow optimized - maybe different Corona settings would affect it (filtering, filter size?).

Corona Filter was set to none to avoid interfering with the filtering process.

[bs]

hi there - is there any news or feature snipped one could miss in regards to the corona denoise feature?

[maru]

*hint* *hint*

https://trello.com/b/EfPE4kPx/corona-road-map-3ds-max

[Ondra]

hi there - is there any news or feature snipped one could miss in regards to the corona denoise feature?

:D

[atelieryork]

Impressive....!

[rozpustelnik]

Finally! Great news! The floor is a bit washed-out from details, though... Will it be possible to set the filtering strenght? How the filtering behaves in animation? Does it work with PT+PT or UHD only?

[racoonart]

Less teasing, more releasing!

[FrostKiwi]

:D

If it's not a flickering mess, then my two cats are getting a corona licence each.
All aboard the Hype train.

[johan belmans]

Less teasing, more releasing!

Yeaah :-)

[tomislavn]

Less teasing, more releasing!

QFT :D

[lacilaci]

can this be used with string options in current daily build, or is it not present at all right now?

[Ondra]

not present at all, sorry

[bs]

love it - thanks for the snipped =)

rendering is starting to make fun again - thanks to corona i must admit

[innobright]

Dear SairesArt,
    Thanks for your thoughts on Altus denoiser and the comparison shots.  We fixed a lot of critical bugs with Altus 1.2 release (Jan 8 2016) and we have quite good results with the filter improvements.  Re: comparison with the magic RPF (random parameter filtering) algorithm, we had this algorithm implemented originally since we have exclusive rights to productize this algorithm (reference to paper  http://www.ece.ucsb.edu/~psen/Papers/Sen11_RandomParameterFiltering_LoRes.pdf).  However, Altus currently deploys RDFC (robust denoising using feature and color information), which performs much better than RPF.  We're fairly confident this would be the best denoising quality you will get, so please try Altus 1.2 and let us know how it works out. 

Regards,
innobright team