Richard Geldreich's Blog: In-loop deblocking of GPU textures

A technical breakdown explains how GPU texture compression block artifacts (seams) are entirely predictable because the block lattice is a known function of UV coordinates and block size, making seam filtering feasible without extra metadata. This applies to formats like ASTC (up to 12x12 blocks) and can be emulated on desktop with BC7 macroblocks. The approach allows in-loop optimization during encoding (using Stochastic Coordinate Descent) that accounts for the seam filter, works with mipmapping and trilinear filtering, and can be applied to normal maps via renormalization. Macroblocks also enable larger DCT/DST transforms for better compression efficiency, with adaptive subdivision for high-detail areas. This is planned groundwork for a future XUBC7 v2 encoder, with ASTC prioritized as the most widely deployed hardware format.

Source: https://richg42.blogspot.com/2026/08/in-loop-deblocking-of-gpu-textures.html. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.