Starting a new thread for this one as it is a specific question. What do you guys think of using DXGI_FORMAT_R10G10B10A2_UNORM for per vertex normals? It has a higher precision than UBYTE4 and one could still use the alpha channel as the sign for a compressed bi-tangent. It looks great to me right now, but I couldn't find many useful references on google, so I guess there is a catch. EDIT: If I doing this stuff right 0.002f results in 0.00293255f after compression to UNORM10 and decompression to float again. Is this plausible? Does it mean that I have to normalize my normals after decompression? That wouldn't be so great anymore. [Edited by - B_old on March 15, 2010 7:00:34 AM]