Topo-DiT: Balanced Routing Bottlenecks for Latent Video Flow Matching

Long video sequences make dense transformer computation expensive. We investigate a latent-grid flow model that aggregates spatiotemporal features into a smaller set of feature slots using balanced entropic transport, processes these slots with a time-conditioned transformer, and broadcasts the resulting features back to the latent grid. The proposed contribution is a testable routing design, not a claim that compact tokens inherently represent physical objects. We specify consistent transport marginals, latent-space training and sampling, an optional masked reconstruction warmup, and matched routing ablations. The implementation passes 26 CPU-only correctness tests without video-model training. No video-generation quality, physical reasoning, memory advantage, or distributed-scaling result is claimed. The current artifact is intended to support subsequent controlled experiments under limited local compute. code: https://github.com/shandingwangyue/topo_dit