Tim Sweeney on Real-Time Radiosity

[XxStratoMasterXx]

I really cant wait to see all the approximations of Radiosity, and then ultimately, arbitrary unconstrained radiosity in games!

Unreal Engine 5 perhaps?

 

[compres]

LOL
Very funny indeed. I would like to add though that Sweeny's engines have always been better in my mind.

I'll take the unreal engine over quake any day, and the last engine from Carmack, is nice on paper, but fails to impress me when I play Doom 3, hopefully it gets better afterwards.

So I call Tim is at the very least on the same league as Carmack. Not below.

 

[XxStratoMasterXx]

LOL
Very funny indeed. I would like to add though that Sweeny's engines have always been better in my mind.

I'll take the unreal engine over quake any day, and the last engine from Carmack, is nice on paper, but fails to impress me when I play Doom 3, hopefully it gets better afterwards.

So I call Tim is at the very least on the same league as Carmack. Not below.

Well the art assets, and minnimum system requirements made Doom not look so hot. However "Engine" and "Game" in terms of "Graphics" are wholly different matters.

To be honest, I think UT2004 looks better than FarCry. Is the engine more sophisitcated? No! It's all in how the developer uses it.

I've found, from a mapping standpoint, Unreal Engine's are very elegant and work very well

 

[compres]

I've always seen better texture quality on unreal engines. Pardon my ignorance, is this what you refer as to mapping?

 

[XxStratoMasterXx]

By mapping, I mean actually making levels

I agree with you...Epic's artists are simply outstanding, the textures and models are absolutely superb!

 

[Reverend]

So I call Tim is at the very least on the same league as Carmack. Not below.

There's mutual professional respect. That was 5 years ago. It hasn't changed as far as I know.

To be honest, I find it ridiculous (and humorous) to read things like "who's a better programmer?".

 

[XxStratoMasterXx]

Hey did you know Carmack spelled Sweeney's name wrong ;-)

 

[Reverend]

Hey did you know Carmack spelled Sweeney's name wrong ;-)

I never would've know if you hadn't told me... oh wow...

 

[AlNom]

*drums rolling* dunh dunh duuuuh.......

 

[suicuique]

Realtime radiosity will eventually be possible. You could implement physically inaccurate single-bounce approximations right now just by precomputing per-object irradiance; 15 years from now there will certainly be enough CPU and GPU power to implement arbitrary unconstrained radiosity. In the meantime, I expect we'll see a variety of approximations emerge.

Can someone explain what he meant by "single-bounce approximations"?

IIRC Radiosity computes the radiation transfer between "patches" (think scene elements e.g. flat surfaces or polygons) of the scene.

The "visibility" of these patches to each other define this transfer (besides the usual parameters like light emittance, reflection factor etc.). And the computation of these "formfactors" (float number in between 0.0 and 1.0 describing how much of the hemisphere of patch A patch B is occupying) is the difficult part.
In case of static geoemtry theses formfactors can be precomputed.
Once i have these all that is to do is to solve a linear equation system defined by a sparsely populated matrix and gueranteeing fast convergence (IIRC).

Where does the "single bounce approximation" come in?
Does he imply that the linear solver just computes the first iteration?

Or does this approach function in a completely different way?

TIA

regards, alex

 

[KimB]

I would expect he's talking about approaching the function in a different way, such as photon mapping or somesuch.

 

[DeanoC]

What Tim probably means by single-bounce approximations is a short-cut to one aspect of GI.

The main visual 'bug' in local light models, is that it doesn't approximate the first reflection of light off a surface. i.e. in reality a white light hits a red wall, other objects will get a fair amount of red light from that 'bounce'.

Its by far the biggest difference between global and local light models. A model that approximates that first bounce, is global in the sense the environment affects the lighting.

There are ways of approximating this first bounce without the doing proper GI. For many surfaces only the first bounce really matters anyway as most energy is absorbed (obviously not true for polished mirror surfaces).

Thats what I expect he means anyway

 

[suicuique]

What Tim probably means by single-bounce approximations is a short-cut to one aspect of GI.

...

There are ways of approximating this first bounce without the doing proper GI. For many surfaces only the first bounce really matters anyway as most energy is absorbed (obviously not true for polished mirror surfaces).

Thats what I expect he means anyway

Ahh. That makes Sense.
But this would not be "real time radiosity" in the traditional sense, would it?

It would be a different approach to model GI style effects (like the "color bleeding" you mentioned).

regards, alex

 

[hovz]

LOL
Very funny indeed. I would like to add though that Sweeny's engines have always been better in my mind.

I'll take the unreal engine over quake any day, and the last engine from Carmack, is nice on paper, but fails to impress me when I play Doom 3, hopefully it gets better afterwards.

So I call Tim is at the very least on the same league as Carmack. Not below.

Well the art assets, and minnimum system requirements made Doom not look so hot. However "Engine" and "Game" in terms of "Graphics" are wholly different matters.

To be honest, I think UT2004 looks better than FarCry. Is the engine more sophisitcated? No! It's all in how the developer uses it.

I've found, from a mapping standpoint, Unreal Engine's are very elegant and work very well

 

[no_way]

There are ways of approximating this first bounce without the doing proper GI. For many surfaces only the first bounce really matters anyway as most energy is absorbed

Hmm .. say you have static lightmaps ready .. what if you render those lightmaps as projected textures on dynamic geometry , projecting it from the original surface ? Wouldnt that be a pretty rough approximation of that first-bounce technique ..

EDIT: i just realized it wouldnt make much sense anyway as the first reason you'd want realtime GI is because of dynamic lights and this would not help in any way

 

[Cryect]

Ahh. That makes Sense.
But this would not be "real time radiosity" in the traditional sense, would it?

Yep, and considering determining patch visibility is the hard part of radiosity there is no reason not to do more than one reflection