Multipass Final Gathering

[wolfinger]

This is simulating Final Gathering with bounces.

I first rendered this picture:

Then "collapsed" it to textures and rendered it again at various settings:

FG Accuracy 1000, Min rad 0.01, Max rad 0.1

FG Accuracy 2000, Min rad 0.01, Max rad 0.1

FG Accuracy 2000, Min rad 0.01, Max rad 100

Notice that while the quality difference between the last two images is almost negligible except for the ceiling the latter one was Significantly faster to render. I think the best results are probably attainable by using a large max radius and more rays to compensate. And from thoughts to actions, here's one more pic:

FG Accuracy 3000, Min rad 0.01, Max rad 100

With some tweaking I'm pretty sure this technique could be used to get very clean results for a cornell-box type scenario, but because this depends on manually making textures from the first pass it's totally impractical for anything except the most simple scene. That is of course until (hopefully) the jolly guys at Softimage and mental images build support for such a feature.

I think that with a good implementation of FG with bounces (and a bit more control in general) we could just about do without GI, and even if not it would certainly add to the options. Caustics would still depend on photons, but that's another story.

Comments?

[gk2k]

decrease your MAX RAD variable significantly. It have to be about 10 * MIN RAD to achive good results. Decrease ACCURACY to about 500 also.

[wolfinger]

I tried that and the image turned out very very noisy, in addition to becoming very slow to render. I think having a higher accuracy and a higher max radius probably has a better quality/time ratio.

[wolfinger]

BTW you'll notice that the second pic in my original post (the first one with the bounce) is rendered with a very small (0.01) min radius and a max radius 10 * min radius, and an accuracy of only 1000. IMO that turned out like total crap, and I'm going to re-render the light textures again with a high max rad and a high accuracy to get smoother results.

[gk2k]

A general rule of thumb to follow is that the Maximum Radius should be about 5%-10% of the scene's size (in SOFTIMAGE units). The Minimum Radius should be about 10% of the Maximum Radius' value.

[wolfinger]

seems like there's a rule for each finger of both hands

does somebody have a (link to a) real good technical explanation of how exactly Final Gathering works. I don't think the couple paragraphs in the manual did a very good job of explaining it.

[atyss]

You should suscribe to the Render List, on this forum.

And the links I gave in the last thread were not enough?

Bernard Lebel

[atyss]

In fact, not the last thread, but the previous one
(Light from an object....)

[wolfinger]

I read the french document, but that was not very helpful. There is not much any information there, but an interesting read nonetheless.

I haven't checked Ed's site yet though, but right now I'm looking for a very detailed explanation on the working of FG. I'll read what Ed has to say tomorrow, but meanwhile I'd still like to see some tech summary or something about FG. Maybe mental images would have something on their website.

Anyway, thanks for the links Bernard! Keep em coming if you have more

[atyss]

Sure

[wolfinger]

I re-rendered the image once I re-rendered the textures with a higher accuracy value. This is the smoothest result I've ever gotten with FG, and it only took 15 minutes to render. The blockiness is only very slightly visible, and there's next to no noise (until you start increasing the contrast in Photoshop of course).

FC Accuracy 3000, Min rad 0.01, Max rad 100

and the same image with a lower Max radius (30 minutes render time):

FC Accuracy 3000, Min rad 0.01, Max rad 0.5

In this last pic the noise is already starting to show, but the lamps and the sphere's shadow looks better. Not worth twice the rendering time though, IMO.

[devdoka]

I think you'll be able to put all this aside in a few weeks.

Mental Images completely rewrote the Final Gathering algorithms for mental ray v3.0. Not only do the images render significantly faster, but they look much smoother with equal or lesser settings.

Matt

---

Matt Lind
Animator / Technical Director
Softimage certified instructor:
Softimage|3D
Softimage|XSI
speye_21@hotmail.com

[wolfinger]

droool

Meanwhile, I got two emails today from this nice lady at Softimage. One contained the XSI 2 keyfile, and the other was malformed! Dammit! I bet that email contained the access info to the XSI 2 site. Now I'll have to wait till next week...

[atyss]

Fess,

I would like you to make a little test. You're right, your image looks great with settings of Ac3000 min0.01 max30.

What I would like to know is if your image still looks clean when a camera is animated in the scene. Is there some noisy flickering?

Thanks
Bernard Lebel

[wolfinger]

hmmm... interesting question! I'll set it up, and let it render tonight, and with luck I can post the animation tomorrow before my big date =)

I have a feeling that the blockiness might be visible in motion, but can't tell without testing.

[wolfinger]

um, just to make sure, which image did you mean? there's one with max rad 100 (the first pic) and onether with max rad 0.5 (the second pic). Neither has a max rad of 30.

The first image took 15 minutes to render, the second takes 30.

[atyss]

Sorry for the bad information.
I meant the one with max radius of 100 (maybe I was distracted by the line below, saying a rendering time of 30 minutes!)

On a side note, may I suggest 2 things, if you have time?
The first is to render the animation wit rebuild photon map on,
The second is to try render again, but with rebuild photon map off.
I'm curious wath would be the results.

Bernard

[wolfinger]

there are no photons here, that's the whole point of this test. I'm only using FG to light the scene.

Rebuilding the textures for animation is such an insanely big task that I won't do it, so any animation will have to be camera-motion only. Because this whole technique is a hackjob of a kind, it's only suitable for the final render of a simple scene with no animation. That is, until Softimage + mental images hopefully build a feature that could do it automatically one way or the other.

[wolfinger]

here's the animation: http://willyt.dyndns.org/f3d/fgtest.avihttp://willyt.dyndns.org/f3d/fgtest.avi

Motion kills the effect unfortunately. The shadow looks just horrible in motion, you can see the blockiness very clearly. Yes, the shadow actually does look like that, it's not because of the DivX compression.

Let's hope mental ray 3 will improve things.