It took only a few seconds to build those large light maps using a high level of details (up to 512 rays) compared to several minutes (or hours) with previous algorithm. Also it performs much better when UV mapping is correctly made (3DMEditor's auto UV mapping features currently don't give great results). This is due to the way it works:
N random directions are generated and for each one the object is rendered to an offscreen buffer using texture coordinates as color components with an orthographic camera. The buffer is read in order to compute the texels visible from infinity. At the end you get an estimation of the amount of light each texel receives from infinity
.It takes almost the same time to compute the light map for small and big models (you can even try with very big models it's still very fast). The only problem is due to rasterization : if a part of a model has a high texture density (higher than pixel density) then its texels may not be detected and this region will be darker than it should.
). Also this full-GPU version doesn't have the problem the other version has : since it does standard shadow mapping occlusion queries for all texels everything that should be lit is list properly 