Velox: Spatial Memory Makes World Models Efficient

What it claims

Velox replaces dense historical KV attention in a chunked video world model with a per-block, per-layer route. It needs no retraining and no change to the generator or its conditioning interface. It is built on LingBot-Fast-Cam. At chunk 48 on a single H200:

• Attention latency drops from about 34.2 ms (dense FlashAttention) to 7.8 ms.
• Denoising time drops from about 7.19 s to 2.96 s.

The authors say the speedup grows with the number of generated chunks. They frame the quality result as a "scoped claim": late-chunk attention cost falls "without materially degrading benchmark quality."

How it works

1. Historical slabs. After each chunk is decoded, a reconstruction backend produces depth, confidence, masks and camera poses. Reliable depth is unprojected and bucketed by global Gilbert-ordered token block. Each block gets a source-camera token rectangle and a low/high percentile depth interval. Validity gates require: - a minimum sample count, - a minimum number of covered tokens, - bounded depth spread, - a minimum slab thickness.

Dense geometry is not kept in the routing score.

1. Query depth by recency-priority multiwarp. The first and last frame depths of recent chunks are cached and forward-warped into the target cameras. Per token, the newest cached frame with at least k_warp samples wins, and its median depth is used. This lets revisits fall back to older observations.
2. Score. Query and historical slabs are approximated as camera-facing oriented bounding boxes (OBBs) in world space. The score is their volume IoU, computed with a separating-axis test followed by a deterministic grid estimate.
3. Coverage gate. A block counts as covered when the fraction of its tokens whose depth falls inside some valid historical slab reaches η_cov.
4. Fusion. Each layer computes its own block importance as a softmax over mean-query · mean-key. In covered blocks, the geometric score softly reweights this importance. Uncovered blocks keep the model's own route unchanged.
5. Safeguards and kernel. The previous chunk is always retained, and global-aligned Gilbert neighborhoods preserve motion continuity. A custom sparse kernel drops whole KV tiles before launch. Fixed blocks avoid the irregular indexing that token-level sparsity would need.

flowchart LR
  H[Historical slab table] --> S[OBB overlap score]
  Q[Multiwarp query depth] --> S
  Q --> C{Block covered?}
  M[Layer mean q·k importance] --> C
  S --> C
  C -- yes --> F[Geometry-reweighted importance]
  C -- no --> U[Model route, no bias]
  F --> R[Route + prev-chunk/Gilbert safeguards]
  U --> R
  R --> K[Sparse tiled attention]

Evidence

• Timing: 48-chunk traces on one H200, with slab construction running asynchronously on the same GPU. Figures are as given above.
• Quality: the 158-case WBench navigation split against dense LingBot-Fast-Cam.
• API-free average: within 0.15 points of dense.
• 19-metric overall average: 75.60 vs 75.52 for dense.

Limits

• The fetched text cuts off partway through the method (Section 3.3). Ablations, per-metric WBench results, hyperparameter values and the stated limitations were not available. Appendix B.3 is referenced but was not included.
• Quality is reported only on WBench navigation, with one base model.
• The approach depends on a streaming reconstruction backend. The fetched text does not name that backend or give its cost, apart from saying it runs asynchronously on the same H200.
• Regions without reliable depth, such as disocclusions or warp holes, get no geometric benefit and fall back to the model's routing. The authors note that stale warped depths can survive disocclusions.
• The text shows no code, weights or licence; it points only to a project page.