arXivDaily arXiv每日学术速递 周一至周五更新

高校专区

Imperial College London(帝国理工学院)

2026-08-31 至 2026-08-31 共收录 2
2605.22967 2026-08-31 cs.LG 版本更新

Learned Relay Representations for Forward-Thinking Discrete Diffusion Models

学习的中继表示用于前向思考的离散扩散模型

Benjamin Rozonoyer, Jacopo Minniti, Dhruvesh Patel, Neil Band, Avishek Joey Bose, Tim G. J. Rudner, Andrew McCallum

机构 * University of Massachusetts Amherst(马萨诸塞大学阿默斯特分校) University of Toronto(多伦多大学) Stanford University(斯坦福大学) Imperial College London(伦敦帝国学院) Mila Vijil

AI总结 提出Learned Relay Representations (Relay)方法,通过可微通道传递潜在信息,使掩码扩散模型在去噪步骤间前向思考,减少推理延迟并提升性能。

Comments 16 pages, 3 figures. Equal contribution: Benjamin Rozonoyer, Jacopo Minniti, and Dhruvesh Patel. Code: this https URL (https://github.com/jacopo-minniti/relay)

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.21366 2026-08-31 cs.CV 版本更新

Relax Forcing: Relaxed KV-Memory for Consistent Long Video Generation

放松强制:用于一致长视频生成的放松KV内存

Zengqun Zhao, Yanzuo Lu, Ziquan Liu, Jifei Song, Jiankang Deng, Ioannis Patras

机构 * Queen Mary University of London(伦敦玛丽女王大学) Imperial College London(伦敦帝国学院)

AI总结 本文提出Relax Forcing机制,通过结构化时间内存提升长视频生成的运动动态和时间一致性,减少注意力开销。

Comments Accepted at BMVC 2026. Project page: see this https URL (https://zengqunzhao.github.io/Relax-Forcing)

详情

展开后加载摘要…

URL PDF HTML 收藏