I’m experimenting with the pictures. messing around with the parameters to DIS flow hasn’t gotten me anywhere nice so far.
for now I’ll revive the initialization from homography and give that a try. local features might still fail. grids are very self-similar and symmetric.
if I’d gotten around to working on that “grid recovery” idea of mine, now would have been a great opportunity to use that.
something dumb like taking center of mass/center of bounding box and orientation based on moments for initial alignment might do well enough.
if all else fails, you could ditch the grid and go with something more textured like Perlin noise or something inspired by it.