The world fastest unbiased renderer!

[exvion]

How I use IGI integrator? What I have to write in SurfaceIntegrator?

[Dimitri]

Nice discussion... One problem however is that nearly all the page of 'Parthenon' is in Japanese... Someone has to be in close collaboration with its author to do something I guess... Here there is another very nice, as it seems, real time try... there is no support for textures in the uploaded version as far as I know but it is open source and the source available in the downloading pages... I do know nothing about programming issues... I justs did think to refer to it here... Its author is very well known, I suspect, from the programming people here... He is the author of Winosi... the first, maybe, serious try to make a usable by everyone unbiased renderer... It never became popular because of it being (as its author openly has declared) very stochastic and solely correctness aimed and because of its author's not having so much time to update it regularly (as again himself says in its site) also...

http://www.winosi.onlinehome.de/Ravi.htm

[PixelOz]

You know, I see a lot of talk about unbiased GPU rendering. Most newer GPU based renderers seem to be based on some sort of unbiased rendering engine of one type or the other and what I would like to know is if unbiased rendering was so slow and now it runs so fast with GPU computing acceleration techniques why don't we see more biased rendering engines being adapted to GPU computing too cause I guess that if you could they would run blazing fast and just like in CPU rendering much faster than the unbiased ones too.

I made a few guesses as to why is this. It could be a number of things, It could be that the appeal of the highly photorealistic capability of such renderers has a lot of allure, it could be that programing wise they are easier to adapt to GPU computing, etc. I just don't know what is the exact reason and I would like to know.

One of the reasons that I bring this up is that not everybody wants photorealistic renders all the time, many times what you want is an efficient render or sometimes you want the look of simple ray traced images. If somebody could show me a renderer that renders regular raytraced images in a biased way very, very fast on a GPU I would be interested in that and I know that many people out there would be interested too cause even if GPU based unbiased renderers run so fast that is still not instantaneous and what many animators sometimes want is not always photorealism, sometimes they want that but in many instances what they want is a fairly good looking render that is fast and efficient and sometimes they want a non-photorealistic rendering of one kind or the other and there are several types.

Back there in one of the posts user tomb said this: "Either way, lux supports both biased and unbiased integrators as plugins so it's quite possible to integrate such techniques if anyone feels the itch" so I wondered if what he meant was that with the future luxrender versions it could do both biased and unbiased renderings that are GPU based. I would like to know if he was referring to that. I would like some input on this from somebody.

I know that the V-Ray people created their new GPU based renderer (V-Ray RT) but I noticed that the new renderer is an unbiased renderer whereas the older renderer is a biased one. What I wonder again is why most companies or open source groups creating GPU based renderers are using unbiased engines instead. Is there an obstacle to create a biased renderer that is GPU based? If so what is the issue? Is it not an obstacle but a decision made based on rendering quality merits? Any other reasons?

I want to make clear that it is not that I don't like unbiased renderers cause I do, it is just that I like some of the characteristic of biased renderers too and personally I would like to have the capability of using both types. And I know that a lot of people out there like some of those characteristics of biased renderers too. For example one of the things that I like about biased renderers is that the rendering time is finite and in unbiased renderings is sort of like it keeps going and refining the image until you are satisfied with the results.

I'm not too clear as how this plays into animation cause in animation you render many, many frames and you cannot render the second until the first is fully finished so for unbiased rendering to be used in animation I guess that you would have to do a few test renders and wait until you get a decent quality and then specify somehow a finite time for each rendered frame based on your tests. I would like to know more about that cause I don't have that much experience along that line. I have rendered many short animations with biased renderers but never with an unbias renderer. I have rendered stills with unbiased renderers though but I have only done a small amount in comparison with other methods.

Sometimes graphic artists like the "computer graphics looks" of renderings because for some illustration work it is very efficient and its "clean" or "regular look" is beautiful for some types of images and also helps sometimes to convey and idea with ease due to it looking so clean and simple. You see this type of illustration in magazines and you see animations rendererd like this for TV documentaries etc. Some of us prefer that type of CG look sometimes instead and I wrote a book recently in which I put this to good use due precisely to its "beautiful simplicity" look and because of its efficiency.

For example sometimes what I want is to take a 3D model and render it with Ambient Oclussion alone or sometimes I want to render it with AO and one or a few other light sources combined but not a fully photorealistic render. Sometimes I would like to use full GI instead to render a still cause it takes longer than with AO but it looks better and different. For some images I want speed and AO is more than enough and for some images I want full GI and I use Blender and in Blender I can have AO with the internal renderer but I can have GI if I use another renderer just like I do with Yafaray. Besides there are other types of non-photorealistic use of renderers today that have a good use like cartoon rendering and many other types and this is still evolving at the moment.

So will we also see other types of renderings that are not photorealistic that are GPU based? Will there be biased GPU renderers in the future? Will there be simpler GPU based renderers that allow you to use AO only when you want to instead of full photon mapping based renderings all the time or no AO at all and just simpler raytracing (and keep in mind that there are different degrees of raytracing quality and features alone)? What are the obstacles to this if any? I would like the more knowledgeable of you out there to give the readers of this forum and me more input on all this.

[tomb]

..So will we also see other types of renderings that are not photorealistic that are GPU based? Will there be biased GPU renderers in the future? Will there be simpler GPU based renderers that allow you to use AO only when you want to instead of full photon mapping based renderings all the time or no AO at all and just simpler raytracing (and keep in mind that there are different degrees of raytracing quality and features alone)? What are the obstacles to this if any? I would like the more knowledgeable of you out there to give the readers of this forum and me more input on all this. ..

Wow, that was a long post Let me just quickly answer this point above - SLG (see the core->gpu forum) already has a "biased" mode (direct lighting), which is basically "classical raytracing" (whitted style), i.e. there are no GI effects at all. It should be very fast indeed. Luxrender also has such an integrator. Secondly, there has been several short animations done with SLG (in full GI path/sppm mode actually) with rendertime per frame in the 30s-2m range, which is not at all bad I think. Also, I believe Octane (a commercial GPU-only renderer) has an AO mode (I seem to have read it in an announcement somewhere). So there you go, dinner is already served

[Down Rodeo]

It might be worth adding that since LuxRays is effectively a method for tracing rays against primitives it could theoretically be used with any render engine, biased or unbiased. I think - might go searching for one of Dade's posts but I'm pretty certain that this was being coded with a nod to portability.

[tomb]

It might be worth adding that since LuxRays is effectively a method for tracing rays against primitives it could theoretically be used with any render engine, biased or unbiased. I think - might go searching for one of Dade's posts but I'm pretty certain that this was being coded with a nod to portability.

Ah yes, that's a good point too!

[PixelOz]

Well all that sounds very good to me.

I just figured you know, that since "regular" or "scanline like" raytracing is so fast already that if it was adapted to GPU computing it would run very, very fast. Also that if AO rendering is already so fast (not as fast as regular raytracing of course but faster than GI and in many cases pretty enough) that it would be very fast if calculated through GPU computing.

For example take a look at this scene of a Star Wars object that I modeled that has close to 283,481 vertexes and 234,091 faces (it is not finished, it has no textures yet cause it is a work in progress):

http://img23.imageshack.us/img23/7920/j ... enotex.jpg

I created and rendered it with Blender 2.49b internal renderer and I rendered it with AO and one regular light and It did it in about 33 minutes in full 1920 x 1080 HD res (the one shown is smaller) with full 16x anti-aliasing but I opened it in the new Blender 2.53 beta and rendered it there and cause the new Blender renderer is much faster it rendered it with those exact settings in about 8 minutes 31 seconds which is a considerable improvement but I figure that it would be way, way faster with AO as it is if it used a GPU based AO renderer and it would be interesting to see how fast it would do it with such a renderer, with time we'll see.

My PC has an Intel i7 920 CPU running at 2.66 Ghz so it has 8 threads due to the HT so it is sort of like a 6 core CPU or equivalent roughly. It has 1300 Mhz DDR-3 memory and I give you that info so you get an idea of what hardware I used to produce those numbers. Now this new PC has two Evga 260 GTX graphic cards and I'm wondering already how much faster I could render that scene or a similar scene with AO if I was able to use the power of those GPUs with AO. Let's see, maybe I'll find out soon enough.

[jeanphi]

Hi,

GPU accelerated biased rendering is what is used for example in games or CAD applications for a few years already. Techniques in there have come along way since first 3D accelerated GPUs and the evolution of OpenGL/DirectX is driven by those needs. One of the most striking papers I've seen recently introduced a very clever technique to render raytraced volumetric effects (including caustics) in real time.

Jeanphi

[Chiaroscuro]

why don't we see more biased rendering engines being adapted to GPU computing too

That's a funny line, discounting what the industry has been doing with GPUs, better and better, all these years now, long before unbiased rendering started to make use of GPUs. I'm still amazed by what "older" games like Crysis (and its Sandbox editor) do in real-time at amazing speeds on current GPUs, and progress hasn't stopped.

[PixelOz]

Yeah but you can't hit the render button in Blender and render with that. You could render the Open GL graphics if you want to but it has no anti-aliasing in the viewports. You can force anti-aliasing through your graphics card control panel but in my Evga SLI cards setup that causes trouble with edge, vertex and face selection so I have to turn it off and some people specially the owners of higher end graphic cards such as Quadro type of cards have managed to anti-alias the Blender viewports graphics sucsesfully but the truth is that many people can't do it without a lot of trouble.

And leave alone textures which appear horribly bad in the viewports and to even get that you have to have UV textures cause procedurals do not show, etc., etc., etc. Those game engine kind of graphics still have tricks and shortcuts used to render them in real-time but these types of engines still have many limitations that regular raytracers do not have.

Even games as impressive as Crysis (and I can play it in HD in my PC with many high quality features on so I know how good it is) that have such good graphics have limitations that regular raytracers can surpass.

Take a look at the example image that I gave you, it has slightly softened reflections in the chrome parts of the model (real reflections) and 3D games do not have that, that stuff is still in prototype stages or experimental but is not production stuff and that image is also rendered with AO and in game engines the AO is pre-rendered as baked and then mixed with the regular textures so to set it up you have to render it first anyway which means back to square one. And again I'm aware of the efforts to create AO or GI that is much faster or real-time but that stuff is still mostly experimental and not production yet.

And in a game engine such as Crysis the textures that produce all those details as the pebbles in the ground for example use some sort of normal mapping or parallax mapping technique but in a 3D modeler such as Blender or any other you can render those pebbles fully round and fully polygonal and it will look better than in Crysis and to be able to do that in a game like Crysis anyway you have to model it highly polygonal in a 3D modeler and then bake that high poly model as a normal map (or the like) and then use that with a lower poly model and as I mentioned before the real high poly model still looks better and that can be used with a regular ray-tracer.

As much as can be done with programmable vertex and pixel shaders and all the tricks that modern game engines use that is still more limited than many modern raytracers. Even Blender's internal rendering engine can surpass most game engines in quality and features and Blender's native raytracer is not exactly known in the industry as the best raytracer but it has improved a lot in the last few years and so have many other raytracers out there. If you don't believe that the Blender internal rendering engine can do that much then go to the Blenders home page gallery and see what it can do even without other raytracers such as Yafaray.

It is not that what you mention is without merits. I am very well aware of what real time game engines can do today so I'm aware of what can be done with real-time graphics and I'm sorry but I believe that stuff was similar to what Nvidia offered with Gelato but Gelato is dead now. I think that it had merits, I believe that its just that it didn't get as far as I think it could have gotten and that it was not well understood by the mainstream at the time so it was superseded by more general GPU computing but most GPU computing at the moment is centering around unbiased renderers and that is just fine but like I said a lot of that stuff is directed toward photorealism and there are many kinds of non-photorealistic renderings that are usable and that have merits that are different to GI photon mapping that people like to do.

Like I said we'll see, I think that we will see much more than just super photon mapping+GI+caustics+radiosity rendering in the near future (after all some of the guys here mentioned that some of that stuff is starting to appear already), I think that it is just that at the moment that is where the main focus is because of the tremendous horsepower gain that GPU computing offers.