Video World Models · 2026-10-16
Velox: Spatial Memory Makes World Models Efficient
Qiuhong Shen, Shizun Wang, Xinchao Wang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Velox is a training-free attention router for chunked video world models. It summarizes historical KV context as per-block camera frustum slabs with depth intervals and uses recency-priority multiwarp to estimate query-block depth. Soft frustum-slab overlap modulates layer-specific block importance, with geometric fallback when coverage is unreliable. It targets late-chunk attention slowdowns in long visual rollouts.
Why it matters
Chunked video diffusion rollouts preserve history via growing KV caches, making later chunks progressively slower even when the camera only sees part of the past. Velox reduces the attention term that grows with rollout length by routing attention at the token-block level using geometry-derived spatial memory, without changing the pretrained video generator or requiring training.
Method
- Training-free inference-time routing: keep the pretrained generator and visibility interface unchanged, but replace dense historical attention with frustum slab routed attention at token-block granularity.
- Represent stored history blocks as camera frustum slabs (source-camera token rectangles + robust depth intervals); score query vs history using world-space OBB frustum slab overlap and coverage from multiwarp query depth.
- Soft geometric fusion: for well-covered blocks, reweight each layer’s model-derived KV-block importance using overlap; uncovered regions fall back to the model-derived route. Add previous-chunk retention and global-aligned Gilbert neighborhood safeguards.
Limitation
Velox is most reliable when the frustum slab table and multiwarp Q-depth explain the current view. Geometry-driven failure modes can occur if reconstructed depth drifts, multiwarp cache fills newly disoccluded tokens with stale far geometry, or OBB approximation overestimates overlap for thin structures; soft fusion can then promote stale or misleading history. Quality tradeoffs include decreased perspective and subject consistency, with dynamics/objects being sensitive to weak geometric correspondence.
Abstract (from arXiv)
Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation. Such interaction requires responsive generation, yet coherent rollouts must retain visual memory of what has already been generated. Chunked video diffusion models usually keep this context in growing KV caches, so later chunks become progressively slower. To address this bottleneck, we introduce Velox, a training-free attention router for region-level context selection using frustum slab correspondences from streaming geometry. Velox represents historical blocks as camera frustum slabs with depth intervals and obtains current-block depth through recency-priority multiwarp. For well-covered regions, frustum slab overlap softly modulates each layer's model-derived block importance. Uncovered regions retain the model-derived route without geometric bias. The fixed token-block granularity matches FlashAttention-style tiled execution and reduces the attention term that grows with rollout length. Frustum slab construction runs asynchronously with denoising on the same H200. At chunk 48, Velox accelerates attention by $4.36\times$ and denoising by $2.43\times$. On the full 158-case WBench navigation split, it remains within 0.15 API-free average points of dense LingBot-Fast-Cam. Please see our project page at https://velox-world.com/velox/ for more details.
Related papers
- World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models
- LOCI: Spatial Linear Memory for Streaming World Models
- Honeycomb: Constant-Size Scene Memory Representation for Video World Models
- Rethinking Video Autoencoders for Interactive World Models
- AgiWorld: An Agentic Video World Model for Building Persistent Worlds
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-15 · Indexed by Pub-AI: 2026-10-16 UTC
Velox: Spatial Memory Makes World Models Efficient
Preprint posted: 2026-10-15 · Original record ↗
- arxiv
- 2610.22311
Explicit source versions (1)
- v1 · 2026-10-15
- Preprint available
Source date: 2026-10-15 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Video world models turn actions and camera motion into long visual rollouts, making video a natural interface for interactive simulation. Such interaction requires responsive generation, yet coherent rollouts must retain visual memory of what has already been generated. Chunked video diffusion models usually keep this context in growing KV caches, so later chunks become progressively slower. To address this bottleneck, we introduce Velox, a training-free attention router for region-level context selection using frustum slab correspondences from streaming geometry. Velox represents historical blocks as camera frustum slabs with depth intervals and obtains current-block depth through recency-priority multiwarp. For well-covered regions, frustum slab overlap softly modulates each layer's model-derived block importance. Uncovered regions retain the model-derived route without geometric bias. The fixed token-block granularity matches FlashAttention-style tiled execution and reduces the attention term that grows with rollout length. Frustum slab construction runs asynchronously with denoising on the same H200. At chunk 48, Velox accelerates attention by $4.36\times$ and denoising by $2.43\times$. On the full 158-case WBench navigation split, it remains within 0.15 API-free average points of dense LingBot-Fast-Cam. Please see our project page at https://velox-world.com/velox/ for more details.
