interesting. i guess the hi normal tangents use 2x16 octo wrap for 4 byte storage? i can not find any other format in the vertex factory shader.
also normal gbuffer shouldn’t need more then 16 bit precision either. 32768 angles is dense enough.
the inbetween 2x12 octo wrap would be interesting for testing and error comparison. it’s got 3 byte storage. but would require shifts and masking cause it doesn’t line up with the channels. hmm…