What is difference between Standard GI and Stochastic GI?

[abdelouahabb]

stochastic is better for animation (no need to calculate GI prepass every time), that's what i found in internet, and for those noise, increase sotchastic samples, and standard GI is better for camera fly for exemple, just an idea...
but in standard GI you've more control because you've more options,
and i dont know if it's the same word but stochastic in probability is like random

[JoelOtron]

While we're on the subject---whats the equivalent terminology for stochastic in other advanced rendering software? Seems in my experience, most non-c4d users aren't familiar with the term.

[rareseu]

if i understand correctly in standard Gi there is number of random sample points that emit stochastic rays, in stochastic mode every pixel emits rays

[Katachi]

Stochastic is inspired by the GI effect introduced by an "old" renderer named Arnold. Afai remember it was the first renderer back then that used a QMC (Quasi Monte Carlo) algorithm for its GI solution and that is now the render engine of Messiah:Render.

But that´s just what is still in my mind. Maybe someone else has even more in depth information on this.

[Per-Anders]

I'm sorry but I would read the manual if I were you, it does explain the GI and it's differences. Stochastic GI is a very simple algorithm, it's a quasi monte carlo sampling for each and every pixel. It produces totally accurate results with some grain and slower render times in most cases. Standard GI samples at fewer points in the scene and interpolates between those samples to create the illumination effect, this creates a generally faster and less grainy but also less accurate GI.

[MJV]

Actually there is no manual, only online help, which for the new user is crippled by the fact that is no mechanism for reading page to page. Even if you find a relevant section using search, the individual pages have no next page previous page buttons, so no way to read your way through as if it were a physical manual. But that's a different issue for a different thread.

As Per noted there is a section in the online help that briefly but competently explains the difference between the different GI rendering options.

http://www.mvpny.com/RadTutMV/RadiosityTut1MV.html

The above tutorial examines in depth the trade off between sampling more or less points in the scene using Standard GI. Once you understand that trade off then it is easy to understand what Stochastic GI will give you by sampling every point in the scene and why it is creates such grainy images.

[he2007]

Thank you for the posts, they've been helpful.

I'm doing more experiments with Stochastic rendering. It seems like Stochastic rendering produces more physically accurate/predictable results than GI standard rendering. I find myself less worried about getting rid of "light splotches" in a scene. Is stochastic rendering considered a type of "unbiased" rendering? Is it more physically accurate in general?

[LucentDreams]

Another thing about stochastic is that while to get comparable rendertimes you generally have a lot of grain, you can also take advantage of Scene motionblur to smooth out the noise allowing you to reduce the samples even more. Generally if I'm animating with cinema's GI its means stochastic. Personally animation without motionblur drives me nuts, so using low stochastic and no AA often counters the rendertime enough to make the speed hit of scene motion blur balance out.

[JoelOtron]

Another thing about stochastic is that while to get comparable rendertimes you generally have a lot of grain, you can also take advantage of Scene motionblur to smooth out the noise allowing you to reduce the samples even more. Generally if I'm animating with cinema's GI its means stochastic. Personally animation without motionblur drives me nuts, so using low stochastic and no AA often counters the rendertime enough to make the speed hit of scene motion blur balance out.

Been using this method like crazy-- one of those hidden gems in c4d as far as Im concerned. The only downside is when you have quick springy secondary motion in your scene. I find that motionblur kind of "stabilizes" the dynamic effect--practically removing it.

[LucentDreams]

I'll cheat my interpolation to favour the keys with less blending to compensate, take an easy ease type curve and pull the handles so far its practically vertical in the middle

[ollle]

Thank you for the posts, they've been helpful.

I'm doing more experiments with Stochastic rendering. It seems like Stochastic rendering produces more physically accurate/predictable results than GI standard rendering. I find myself less worried about getting rid of "light splotches" in a scene. Is stochastic rendering considered a type of "unbiased" rendering? Is it more physically accurate in general?
I wouldn't call it "unbiased". But similar to the "new" unbiased renderers (like Fry or Maxwell) the Stochastic renderer calculates the GI for every pixel.
If you set "Stochastic Samples" (and "Diffuse Depth" of course) to very high values it gets very close to unbiased.

[abdelouahabb]

so, if the stochastic dont use interpolation (the grainy image), so no need to adjust the illumination channel in the material propriety?
but i think standard gi is better, because you can save your gi solution, and i think you can save it for a 640*480 and re-use it on 1024*768?
and you can decide how you will sample the scene, the prepass 1/1 means that you sample every pixel in the "snapshot", and so on....sorry if i'm false but that's what i understand :)

[JoelOtron]

I'll cheat my interpolation to favour the keys with less blending to compensate, take an easy ease type curve and pull the handles so far its practically vertical in the middle

Thanks for the tip

[Ernest Burden]

I'll cheat my interpolation to favour the keys with less blending to compensate, take an easy ease type curve and pull the handles so far its practically vertical in the middle

Kai

I've experimented with your suggestion in the past but not 'gotten' it. I'm one of those people who animate without motion blur (camera animation usually, arch-vis stuff) so I'm not familiar with the settings at all. Would you be kind enough to post a sample file with the setting you have described for the stochastic w/mo-blur technique?

For those interested in NPR, one of the first things I did when I bought C4D years ago was discover that stochastic at super-low settings produces a look very similar to a Black Prismacolor drawing.

[georgedrakakis]

personally i like grain, especially in clay-look models,
and yes, i agree with Ernest, ultra low settings produce far more interesting results
http://i122.photobucket.com/albums/o271/studiomoccoro/opsi_22.jpg

[stew]

I wouldn't call it "unbiased".
Why not? For all I know, it is unbiased QMC path tracing.

[christianS]

Why not? For all I know, it is unbiased QMC path tracing.

Well, i guess it's because Maxwell and the like (basically all MLT-Renderers) were the first to use the word as some kind of new must have feature. Now it is somehow associated with those Renderers and their physically correct results although it is quite a bit older. Its somehow like the old mixing of Radiosity and GI. I would also think "standard" is "biased" and "Stochastic" is unbiased.

[stew]

Yes, where I don't know why they're emphasizing the unbiased part so much, since it's neither new nor interesting. An average CS student who took a computer graphics course can write an unbiased path tracer. The MLT part and spectral rendering is what they have and the others don't (along with fine-tuned material and light shaders), but being unbiased is nothing to brag about. In fact, step 1 of writing a biased GI renderer is to write an unbiased one...

http://renderspud.blogspot.com/2006/10/biased-vs-unbiased-rendering.html
http://www.cgafaq.info/wiki/Bias_in_rendering

So yes, while it doesn't explicitly use those terms in the UI, C4D's stochastic GI is most likely a QMC path tracer and the non-stochastic mode is obviously using irradiance caching. I doubt there's much hope that one day people would use the correct terms for things...

[JoelOtron]

Kai

I've experimented with your suggestion in the past but not 'gotten' it. I'm one of those people who animate without motion blur (camera animation usually, arch-vis stuff) so I'm not familiar with the settings at all. Would you be kind enough to post a sample file with the setting you have described for the stochastic w/mo-blur technique?

For those interested in NPR, one of the first things I did when I bought C4D years ago was discover that stochastic at super-low settings produces a look very similar to a Black Prismacolor drawing.

In case kai isnt around...

There was a great post by Simon Wicker a couple of years ago, where I first learned of this technique. Cant seem to find it though...

Basically, pretty straightforward--turn on stochastic GI, turn off AA, and then turn on SCENE motion blur in your post effects.You will then need to fiddle with the settings of GI and moblur to get the right speed/quality balance. You might be able to get away with turning down your accuracy (so there wouuld be fewer passes per frame) and samples down quite a bit, so you are left with a very noisy render. Then adjust your moblur--increasing the samples and strength as needed. The motion blur not only blends and smooths out the noise as Kai mentions, but is also bypassing the need for AA altogether (something I didnt figure till about 2 years ago!) as each frame is smoothed out. You may run into issues when using transparency and reflection, which may lead to a loss in reflection/transparency detail or increasing render time dramatically. Its the kind of thing that will work well in certain instances but maybe not so well in others. I guess try a frame or 2 from an already lit scene file.

Im sure if kai returns to the forum he might be able to give more info.

[georgedrakakis]

In case kai isnt around...

There was a great post by Simon Wicker a couple of years ago, where I first learned of this technique. Cant seem to find it though...

i think it's located here:
stochastic GI tip (http://forums.cgsociety.org/showthread.php?f=95&t=425605)
great stuff, i'll surely give it a try.
cheers,
george

[JoelOtron]

i think it's located here:
stochastic GI tip (http://forums.cgsociety.org/showthread.php?f=95&t=425605)
great stuff, i'll surely give it a try.
cheers,
george


Good find!
Ha--Im in that thread too--guess I lost the bookmark.

[LucentDreams]

Okay setting wise in scene motionblur.

Strength 0-400% This setting basically controls the favoring from one key tot he other, but the hey thing to remember is the first pose of a key is always counted. So 1 or 2 % will make sure that your basiclaly just blurring the frame itself with very little motion, 400% will favour the second frame but will always have a strong ghost of the first frame so usually not usable. I usually go from 2%-35% for most work. Usually if banding is showing up I reduce the strength before I resort to using more samples. Depending on the type of motion I try not to go to 25 because of the substantial render addition so I'd rather try and make 16 work if possible, but for slow character stuff I've even used 9 before in 2K feature film. For the Stochastic GI or AO trick though keep in mind fewer GI/AO samples means more motionblur passes to help counter the grain.

Dither o -100% This basically dithers between each of the x number of passes which helps to hide the banding of all the passes. That said don't just assume if theres banding that setting it to 100% will solve things. Typically 100% will create more grain and reduce the overall visibility of the blur. I usually use 25 a max of 75 I generally don't use 0 at all if I don't need much I set it to 5% just to break up any edges.

Anti aliasing restriction, This turns off AA internally for anything about 9 samples I believe, maybe its 16, I'd have to check the manual for sure, but I typically turn off the AA manually anyways.

Camera offset, never turn this off imo. Especially for the stochastic/AO trick it won't really work without this. It basically vibrates the camera a little each sub frame to get a more random sampling instead of straight linear.



On the stochastic side. Generally just us a diffuse depth of 1,m 2 at max, remember it doesn't have the same sort of color bleeding that standard GI has. The samples is the Max number of samples. there is a range believe it or not but you set it using accurracy. at 25% its going to favor less samples and at 75% it will favor more. typically I'll set the max number a little bigger than my Ideal and drop accuracy down, that way picky areas will get a little more detail while still keeping overall time lower. Finally note the identical noise distribution is still on, TURN IT OFF. The effect doesn't really work well if its on. I think simon said this in his too. Its a simple logic though if the samples are always identical then the blurring is only the subtle camera offset so you going to have larger blotches for each noise grain. With it off the samples are random every single sub frame and the blurring of all the passes will create a smoother wash.

[ollle]

Why not? For all I know, it is unbiased QMC path tracing.
My defintion of unbiased: A brute force kind of rendering without adaptive techniques, that gets the most correct solution at the end. I would call the stochastic mode unbiased, when you set the values for "stochastic samples" and "diffuse depth" to infinite (but that's not possible).
But that's only my personal definition :-)

[stew]

My defintion of unbiased: A brute force kind of rendering without adaptive techniques, that gets the most correct solution at the end.
That attribute is commonly known as consistent (http://www.cgafaq.info/wiki/Bias_in_rendering). I know that the Maxwell manual say something different, but it's wrong. Given enough render time, photon mapping will also converge to the correct result.

[Ernest Burden]

Found it...this is what I was talking about:

http://www.oreally.com/temp2/C4-test02.jpg

Super-low samples and pumped-up GI strength

[CGTalk Moderation]

This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.