The Pulse
Velox Cuts Long-Run Video Attention Time 4.36×
A National University of Singapore team says its Velox method cuts video world-model attention latency by 4.36× at chunk 48. The training-free router stays within 0.15 points of dense attention on the paper’s 158-case WBench navigation eval
AI.info Team · Written with AI ·
At chunk 48, Velox cuts a video world model’s attention latency from 34.2 milliseconds to 7.8 milliseconds, a 4.36× reduction, according to a preprint submitted to arXiv on October 15. The method also cuts denoising time for that chunk from 7.19 seconds to 2.96 seconds. Those are late-rollout measurements: Velox’s speed advantage grows as a model generates more video and accumulates a longer history.
Velox routes attention by camera view
Researchers Qiuhong Shen, Shizun Wang and Xinchao Wang, at the National University of Singapore, target a cost that builds into chunked video generation. Each new chunk can attend to the growing cache of earlier video, even when the current camera view overlaps only a small part of that history. Velox keeps the pretrained generator and its conditioning interface unchanged, altering which historical blocks are visible during inference.
The router represents historical video blocks as camera-frustum slabs with depth intervals. It estimates depth for the current view, then uses geometric overlap to favor history that corresponds to the regions being generated. Where the geometry does not adequately cover a region, the method falls back to the model’s own attention-based route; it also retains the previous chunk and nearby blocks as safeguards.
“At chunk 48, Velox accelerates attention by 4.36× and denoising by 2.43×.”
Qiuhong Shen, Shizun Wang and Xinchao Wang, “Velox: Spatial Memory Makes World Models Efficient”
The speedup arrives late in the rollout
The reported timing comes from a 48-chunk rollout on one Nvidia H200, with construction of the spatial geometry running asynchronously on that same GPU. The paper says dense attention is faster during the first few chunks, and the Velox and dense timing curves cross within the first seven. The 4.36× figure therefore describes attention at chunk 48, not a uniform speedup from the start of generation.
Denoising at chunk 48 still takes 2.96 seconds in the reported setup. That is substantially less than the dense-attention result, but it does not by itself demonstrate real-time interactive generation. The test measures the efficiency of a late chunk under a specific GPU setup, rather than a full system’s response time in a live simulation.
WBench scores stay close to dense attention
The team evaluates quality on all 158 cases in WBench’s navigation split, using LingBot-Fast-Cam as the dense-attention baseline. Velox scores 73.51 on the paper’s API-free average, compared with 73.66 for dense attention—a 0.15-point gap. On the separate 19-metric overall average, Velox scores 75.60 against 75.52 for dense attention.
Those figures support a narrower claim than a general guarantee of unchanged video quality: on this benchmark and model setup, the authors report near-parity in aggregate scores while reducing late-chunk costs. The paper also compares against a model-only block-sparse route, which scores 73.11 on the API-free average. Velox’s geometric routing raises that result to 73.51, according to the study.
Code is still listed as planned
Velox is training-free, but its implementation still depends on the paper’s block-sparse attention route, geometric memory construction and asynchronous execution. The project page says code and results are planned, so independent users cannot yet reproduce the reported measurements from a public release listed there. The next practical test is whether the speed and quality results hold when the method moves beyond this 48-chunk evaluation and its single-H200 setup.