NVIDIA / MIT reconstruction paper

[ktegan]

This paper looks really interesting. They get some pretty impressive results with 16 samples / pixel, and their reconstruction times are in the order of 10 seconds using a GPU. View dependent effects (such as glossy highlights) seem to be one area where this might have some problems (or possibly not as many benefits).

Temporal Light Field Reconstruction for Rendering Distribution Effects
http://www.tml.tkk.fi/~samuli/

Kevin

[jeanphi]

Hi,

The memory usage is huge because you need to keep track of all samples. I also don't know how it would react to MLT algorithms.
This other paper from the same page looks interesting too: http://www.tml.tkk.fi/~samuli/publicati ... _paper.pdf

Jeanphi

[ktegan]

Hmmm... it seems like there are two possible avenues for reducing memory. Rendering and reconstructing in tiles is one solution. Another interesting solution is trying to aggregate samples that have similar motion / depth / lens parameters.

I guess primary visibility effects such as this are often resolved well before the global illumination, so this paper may not have a huge value for typical lux renders.

Kevin