Quick-'n-Dirty Radiosity

quote:

---

I mean, radiosity that''s quick, easy, and preferable to typical stencil shadow volumes?

---

Ehm, no. Quality radiosity isnt quick. You can fake it to be near real-time (for simple scenes) but it definitely isnt real-time alternative to stencil shadow volumes.
Google it or just go here for some basic info on fast radiosity: http://freespace.virgin.net/hugo.elias/radiosity/radiosity.htm
Basically you divide the model into patches (triangles or quads) and let every models patch be a light source radiating light to every other patch in a model. You have to stop the algorithm at some point and it depends on you whether you want to wait long for quality result or faster less-quality one.

While I do not know what specific radiosity algorithm does he use (he implemented most of them, AFAIK), I think that he means just radiosity that converges fast and doesnt need to be precisely calculated on supercomputers (which costs hell of a money).

VladR
Avenger game

Well, the word ''quick'' is highly relative... I meant quick in terms of fast convergence and low-quality, since I didn''t want to wait hours for some test trees.

The radiosity I used is a basic progressive refinement system, taking only vertex colour into account. That''s usually enough for a tree, considering the high tesselation. I first shot the primary light from a HDRI envmap, but only monochromatic, distributed over the full spectrum (ie. greyscale lighting). I then do a quick PR convergence over the tree mesh, by path tracing only, no hemicubes. Taking into account the leaf texture alpha, of course. This step is multi-spectral, ie. takes into account the full colour spectrum, to allow colour bleeding, although I started with a monochromatic primary light. That''s because I wanted the tree-inherent colour bleeding (the green leaves, etc), but no influence from an external spectrum. Not at this point.

So I had a tree, with per-vertex radiosity (RGB) that would represent the internal lighting, assuming a monochrome external spectrum. The realtime lighting is done via a combination of a local per-tree incident illumination cubemap (indexed via per-pixel EMBM), combined with the per-vertex radiosity. The illumination map is computed at scene build time, by using the full radiosity solution for the entire 3D world.

That way, I can instance a single tree geometry, but still have full local lighting (with soft shadows). And all I have to store per instance, is a small 16*16*6 cubemap. That''s not even 5kB per tree or tree group, however complex the reference tree model is. The internal per-vertex tree radiosity solution is the same for all trees.

So in short:

* internal self shadowing and colour bleeding is done once per tree mesh, through per vertex PR radiosity.
* external radiosity and soft shadows are added per-instance, by using an illumination incidence EMBM cubemap.
* Sharp shadows and local lights are added by shadowmaps and dot3 lighting, per-instance.