@k_narkowicz @aras There are two main variants: one is storing the ZH3 coefficient (so an extra RGB triple, going from 13 -> 16 bytes with the ratio-based encoding), and the other is hallucinating the coefficient based on the L1/L0 ratio (so no extra storage but a little extra computation). Explicitly stored is the best quality (caveats about interpolation aside), but the hallucinate (which is what’s in the teaser image) still is better than linear SH in the vast majority of cases.