I used a similar method of quantizing normal vectors, but I settled for 240 of them, so one byte per normal. It was based on IIRC using a dodecahedron, split fa

I used a similar method of quantizing normal vectors, but I settled for 240 of them, so one byte per normal. It was based on IIRC using a dodecahedron, split faces into triangles, then split those each into 4. Then use the face normals. I dont recall it being as complicated as it sounds, but it does work rather nicely. I could precompute view dependent shading for each normal once per frame and then use the normal index as a color index instead. CPU implementation.