AnchorWeave World-Consistent Video Generation with Retrieved Local Spatial Memories paper:

The WASD overlay really sells the "playable" aspect here. Smart move by the UNC team to ditch global 3D reconstruction for those local geometric memories—seems like a much cleaner way to handle drift without the usual misalignment noise.

I personally think that explicitly modeling 3D structure directly in 3D is difficult to scale, so it doesn’t really seem worthwhile.

We approach temporal consistency via slot attention with Hungarian matching - cosine sim jumps from 0.365 to 0.907 across frames. Curious how their spatial memories compare to learned slot correspondences on longer sequences!