Research map / Video / Video generation
AdaState: Self-Evolving Anchors for Streaming Video Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Replaces the static first-frame KV anchor in streaming autoregressive video diffusion with a denoised hidden "adaptive state" latent carried via the KV cache, trained with horizon-weighted DMD, to improve dynamics without sacrificing consistency.
Problem
Autoregressive streaming video diffusion models anchor to a static first-frame KV, which suppresses motion and scene progression in favor of static consistency, producing temporally shallow videos especially beyond the training horizon.
Approach
A hidden adaptive state latent is denoised jointly with content at every chunk (never rendered), its clean KV overwrites the position-0 anchor for the next chunk, making denoising act as a recurrent state transition over the KV cache; training uses a linear per-frame horizon weighting of the DMD loss.
Key results
- At 5s, AdaState reaches VBench total 0.860 and dynamic degree 0.828, the highest among compared methods (Self-Forcing 0.832 total, 0.635 dynamics).
- At 30s (6x training horizon), AdaState achieves VBench total 0.865 and dynamic degree 0.922 while keeping consistency at 0.259, the best among methods; MemRoPE is second at 0.848 total.
- A user study with 40 Prolific raters ranks AdaState highest on coherent progression and prompt adherence against four baselines.
Surprising findings
- Retaining a static sink alongside the adaptive state suppresses dynamics (Sink + State dynamics 0.570 vs 0.828 without sink), indicating the model shortcuts through the clean static reference.
- Adding more state slots (F_s=3) does not improve dynamics; a single state slot in a tighter window performs best.
Limitations
- State capacity fixed heuristically at one frame; longer rollouts or complex multi-actor scenes may need more state or explicit memory.
- Subject consistency is slightly lower than static-anchor methods (0.961 vs ~0.98 at 5s), the cost of an evolving reference.
- Evaluation focused on video; generality to other autoregressive domains is untested.
Experimental setup
Datasets: MovieGenBench Self-Forcing training prompts
Benchmarks: VBench VisionReward MovieGenBench
Metrics: VBench total subject consistency dynamic degree aesthetic quality overall consistency VisionReward background consistency motion smoothness imaging quality temporal flickering
Baselines: Self-Forcing CausVid Causal Forcing LongLive Rolling Forcing Infinity-RoPE Reward Forcing MemRoPE Rolling Sink Wan 2.1-1.3B
Builds on: Self-Forcing Distribution Matching Distillation Wan2.1 Infinity-RoPE block-relativistic RoPE Causal Forcing
Closest papers by meaning
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.90)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.90)
- Flex-Forcing: Towards a Unified Autoregressive and Bidirectional Video Diffusion Model (ICML 2026 · similarity 0.90)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.90)
- Streaming Autoregressive Video Generation via Diagonal Distillation (ICLR 2026 · similarity 0.90)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.