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.
By Zengqun Zhao, Yanzuo Lu, Ziquan Liu, Jifei Song, Jiankang Deng, Ioannis Patras
arXiv:2609.35491v2 Announce Type: replace-cross
Abstract: Few-step autoregressive video generation commonly relies on Distribution Matching Distillation (DMD), requiring a bidirectional diffusion tea...
By Chi Zhang, Yueyi Liu, Haoyang Shi, Ruichuan An, Haoyu Li, Yuhang Wu, Sen Cui, Miao Liu
arXiv:2610.10429v1 Announce Type: new
Abstract: Autoregressive video generation requires denoising the current frames while writing their key-value representations as context for future predictions....
By Zihan Su, Junhao Zhuang, Yaowei Li, Siwen Lu, Haoran Li, Lingen Li, Haoyu Wu, Weiyang Jin, Songchun Zhang, Haoyang Huang, Chun Yuan, Zeyue Xue, Nan Duan
arXiv:2607. 20125v1 Announce Type: cross Abstract: Autoregressive (AR) video diffusion models have become a promising paradigm for long and streaming video synthesis, but the continuously growing Key-Value (KV) cache makes attention the dominant inference cost, especially at high resolution where each frame contributes many tokens.
By Jinliang Shen, Lianghao Su, Zheming Li, Kang He, ZiLiang Lai, Yanbing Jiang, Chengru Song
The paper introduces Recency Forcing, a technique that addresses the long‑horizon degradation in autoregressive video generation caused by KV eviction mismatch. By applying a timestep‑dependent bias—Temporal Response Bias—derived from a positional response measure, the method reduces the influence of distant frames during inference without altering context length or training objectives. An exact reformulation, Biased Attention Reparameterization, enables this bias to be applied as a standard FlashAttention call with zero overhead, achieving state‑of‑the‑art long‑horizon generation quality on VBench datasets.
By Tri Cao, Hung Nguyen, Phong Nguyen, Khoi Nguyen
In-Distribution Forcing for Long Video Generation at Test Time introduces a test‑time framework that aligns KV caching and conditioning with training configurations to prevent out‑of‑distribution key‑value entries during long‑horizon video generation. The core technique, self‑caching, ensures each chunk is cached without attending to prior KV entries, keeping the rolling window exactly in‑distribution. This approach extends short‑horizon autoregressive video diffusion models to minute‑scale generation, outperforming prior methods in mitigating drifting as shown by drift metrics and a user study.
By Jeongwoo Shin, Youngyoon Choi, Sangwoo Jo, Hyunmog Kim, Sungjoon Choi, Joonseok Lee, Jaewoong Choi, Jaemoo Choi
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.
By Yu Meng, Xiangyang Luo, Letian Li, Wenyuan Jiang, Chen Gao, Xinlei Chen, Yong Li, Xiao-Ping Zhang
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.
arXiv:2605. 19398v3 Announce Type: replace-cross Abstract: Image-to-video models often generate videos that remain overly static, compared to text-to-video models.
By Wooseok Jeon, Seungho Park, Seunghyun Shin, Sangeyl Lee, Hyeonho Jeong, Hae-Gon Jeon
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.
By Mahmoud Ahmed, Sameh Abdulah, Olatunji Ruwase, Sam Ade Jacobs, Mathis Bode, Mohamed Elhoseiny
arXiv:2605. 21028v2 Announce Type: replace-cross Abstract: Autoregressive long video generation often adopts bounded-memory streaming for efficiency, typically combining local windows for short-term continuity with static early-frame sinks as long-range anchors.
By Bo Ye, Xinyu Cui, Jian Zhao, Tong Wei, Min-Ling Zhang
The paper introduces Connected Self Forcing, a training framework that reconnects gradient paths across autoregressive video chunks to improve long‑horizon visual quality and temporal consistency. It allows feedback from later predictions to influence earlier context generation, using shortcut gradient replay to keep memory usage manageable. The method is combined with distribution‑matching distillation and shows experimental gains without altering the inference procedure.
By Dongbin Zhang, Chaoda Zheng, Kangjie Chen, Xiangyu Li, Shijia Chen, Jinhao Deng, Yuqi Zhang, Guangfeng Jiang, Hongbin Lin, Choo Sin Wai, Minqi Wang, Puyi Wang, Jingye Zhang, Yu Zhang, Xianming Liu, Boyang Wang