I'm talking about estimating the direction of padded pixels in the correct tangent space, not blending to the other side edge color. It doesn't even have to be necesseraly estimated as it could be known. No ?
@poopipe Did not try pixel snapping, just tried and the problem persists, marginally better than the above screenshot for SP2 bakeLike so: @Obscura Tried various dilatations from 4 to 16 to maximum, still the same result, and in general i tend to always put at least 4-8px,Something i just remember reading somewhere is that…
Well, maybe, but I'd be very curious to see any case where such a convoluted post-process (involving somehow fetching the normals information from the "other side" of an edge, converting back to the proper tangent space of the initial side, and feeding that into the padding of an island edge) would have any impact. All…