The paper introduces Relax Forcing, a training‑free memory mechanism for autoregressive video diffusion that structures temporal context into Sink, Tail, and History frames. By selecting History frames with a relaxation criterion, the method reduces error accumulation and attention overhead while preserving motion dynamics. Experiments on VBench‑Long demonstrate that this structured memory improves long‑video generation quality over existing baselines.
arXiv:2606. 14732v1 Announce Type: cross Abstract: Autoregressive video diffusion models enable streaming generation but often degrade over long rollouts: static scene layouts drift, while mechanisms that improve spatial stability tend to suppress motion, causing natural flows such as water, fire, or smoke to stagnate.
Ring Forcing is an autoregressive video diffusion framework that enhances long‑term memory by enforcing retrieval from distant history through a ring‑structured training strategy. It introduces a compression and timestep composition method to extend effective historical span to minutes, and a sparse RoPE mechanism for scalable memory adaptation. Experiments show that Ring Forcing outperforms state‑of‑the‑art models in minutes‑long coherence and object permanence.
LiveVVT introduces a rolling streaming diffusion framework for video virtual try‑on that maintains high visual fidelity while enabling real‑time performance. It preserves bounded bidirectional modeling within a fixed‑size window, emits clean video chunks iteratively, and uses two memory modules—a bounded temporal memory and a persistent global appearance memory—to sustain long‑term consistency. A progressive distillation process further aligns teacher‑based bidirectional learning with causal few‑step inference, resulting in superior generation quality with 26× lower latency and 11× higher throughput compared to comparable 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:2607. 15271v1 Announce Type: cross Abstract: Online novel view synthesis from multi-view streaming videos faces a fundamental trade-off: maintaining a persistent, long-horizon memory to reconstruct temporarily occluded regions while operating under strict real-time constraints.
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.
LiveVVT introduces a rolling streaming diffusion framework for video virtual try‑on that maintains high visual fidelity while enabling real‑time performance. By confining bidirectional spatio‑temporal modeling to a fixed‑size window and using bounded temporal and global appearance memories, it emits clean video chunks with low latency. A progressive distillation pipeline further refines the model, achieving superior quality with 26× lower latency and 11× higher throughput compared to prior methods.
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 have emerged as a promising approach for long video generation, achieving strong performance in streaming settings. However, existing methods are restricted to forward temporal generation, whereas practical video creation often requires flexible generation order, e.
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.
LongVU‑TTT is a causal test‑time training method for long‑video multimodal large language models that inserts a convolutional resampler with fast‑weight updates between the vision encoder and the LLM. The fast weights adapt per video and contextualize frame features before compression, while a hybrid selector keeps explicit visual evidence for downstream reasoning. Experiments show that TTT‑Conv outperforms TTT‑MLP and bidirectional Mamba2 on MLVU, and beats attention‑ and fixed‑state recurrent resamplers on three benchmarks, achieving competitive results on five video‑understanding tasks after reducing 512 frames to 128 LLM frames.
arXiv:2608. 07408v1 Announce Type: cross Abstract: We study visual persistence in interactive video world models.