Paper Overview
Research Area: Computer Vision (CV) Authors: Yusuf Dalva, Pinar Yanardag Published: 2026-05-28 arXiv: 2605.30349
Summary
Autoregressive video diffusion models enable streaming video by generating frame by frame, with each segment conditioned on previously generated content. These models are structurally anchored to the first frame: its key-value representations occupy a privileged position in the attention cache and serve as the primary scene reference throughout generation.
As the cleanest, least erroneous position in the cache, this anchor attracts disproportionate attention, which suppresses video dynamics and locks the scene composition to the initial viewpoint—even as the scene naturally evolves. The result is temporally shallow video, where motion, camera movement, and scene progression are suppressed by static consistency.
The Proposed Approach
The authors replace the static anchor with an adaptive state—a hidden latent variable that the model processes in parallel with each segment's denoising but never renders. Instead of referencing a frozen first frame, the model generates its own scene anchor at each step by attending to the previous state and the current content, producing a reference that evolves with the generated content.
Unlike standard video generation, which encodes absolute temporal concepts, this formulation treats time as relative: every generation step sees the same pose structure regardless of how far generation has progressed, and the state transition is identical for every segment. Together, these properties introduce recursion into the generation process, where denoising acts as the transition function and the KV cache serves as the carrier—without any external modules.
Results
Experiments demonstrate that adaptive states significantly improve video dynamics, enabling richer motion and more natural scene progression.
Links
- arXiv: https://arxiv.org/abs/2605.30349