Addressable Memory for Video World Models
arXiv:2608. 07408v1 Announce Type: cross Abstract: We study visual persistence in interactive video world models.
arXiv:2608. 07408v1 Announce Type: cross Abstract: We study visual persistence in interactive video world models.
LayerRecall is a memory router for autoregressive video diffusion that selectively retrieves and injects historical key/value states into specific layers of the model, based on the current context. It addresses the problem that existing memory mechanisms expose nonlocal history but do not guarantee effective use, by recognizing that different layers prefer current, recent, or distant context. The method, combined with Cross‑Horizon Prediction Matching, achieves state‑of‑the‑art long‑range consistency on MemoBench and MovieBench while maintaining local continuity and incurring negligible inference overhead.
arXiv:2606. 13035v1 Announce Type: cross Abstract: Autoregressive video diffusion models provide a natural formulation for streaming and variable-length video generation by conditioning newly generated frames on previously generated content.
Autoregressive video diffusion models provide a natural formulation for streaming and variable-length video generation by conditioning newly generated frames on previously generated content. However, extending these models to minute-level generation remains challenging: the limited KV-cache budget prevents the model from retaining the full history, while repeatedly conditioning on self-generated frames induces a context distribution shift that accumulates over time, leading to visual artifacts, quality degradation, and temporal drift.
The paper introduces Learning-to-Retrieve (L2R), a method that embeds memory retrieval directly into video world models. By using the model’s internal state as a memory and employing a camera‑conditioned retrieval gate and a retrieval trigger supervised by a 3D re‑visibility signal, the approach learns when and what historical information to access. This eliminates the need for external memory systems and improves long‑term scene consistency across various base models and camera‑revisit benchmarks.
The paper introduces TetherMem, a training‑free, query‑aware memory router designed for streaming autoregressive video models that generate long videos in chunks. By separating subject and scene queries and modulating historical access with region‑ and age‑conditioned priors, TetherMem prevents the model from anchoring the scene to stale backgrounds and viewpoints, a problem termed memory‑anchored scene under‑progression. In blinded pairwise evaluations, TetherMem outperforms eight baseline methods in overall quality and scene progression, and on full 30‑second videos it maintains background, viewpoint, and scene changes while preserving subject identity and temporal continuity.
The paper introduces a decoupled memory control framework for spatially consistent long‑horizon video generation. By separating memory conditioning from generation, it uses a lightweight memory branch to learn spatial consistency from past observations and a per‑frame cross‑attention mechanism to condition each frame on the most relevant historical data. A camera‑aware gating mechanism further ensures memory is used only when meaningful references exist, resulting in lower training costs, higher spatial consistency, and state‑of‑the‑art visual quality.
RECAP-Forcing is a new method for long autoregressive video generation that addresses the memory challenge by organizing memory based on appearance novelty rather than recency. The approach retains key-value caches for newly appearing content—such as entering subjects, disoccluded regions, and new scenes—at the moment they first appear, ensuring consistent identities over time. It combines an attention sink for the initial scene with an optical-flow-based novelty bank for later frames, improving visual quality and semantic fidelity without adding learnable parameters.
arXiv:2610.01192v1 Announce Type: new Abstract: Streaming vision-language models must process continuously growing video streams under a bounded compute budget, creating a persistent tension between...
Momentum introduces a two-stage pipeline for video world models that uses synchronized multi-camera rigs to represent evolving scenes. The WorldUpdate stage generates shared reference-camera videos as a memory of scene dynamics, while WorldInteract grounds target camera video generation in this memory, ensuring moving objects remain represented outside the field of view. Experiments on CARLA street scenes show a 126.2% improvement in dynamic scene consistency over static image conditioning.
arXiv:2606. 09803v1 Announce Type: cross Abstract: We present \textbf{Echo-Memory}, a controlled study of memory mechanisms in action-conditioned world models.
We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout.