Research map / Video / Video generation
Disen-Forcing: Disentangling Semantic Anchoring from Motion for Autoregressive Video Diffusion
Zhengyang Yu, Akio Hayakawa, Masato Ishii, Qingtao Yu, Xuesong Li, Takashi Shibuya, Yuki Mitsufuji, Jing Zhang
Read the paperCheck your idea against 137,217 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.
Proposes Disen-Forcing, a method for stabilizing long-horizon autoregressive video diffusion rollouts by disentangling semantic anchoring from motion, building on frame sink ideas.
Problem
Autoregressive video diffusion models degrade rapidly when the inference horizon exceeds the training horizon, due to compounding exposure bias in long rollouts.
Approach
Revisits the heuristic frame sink design (persistent KV cache of initial frames, inspired by attention sinks in streaming LLMs) and separates semantic anchoring from motion dynamics to stabilize long generation. Full details are not available from the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add disentangled context routing to Frame Sink baseline (ablation) | Self-Forcing + Frame Sink · MovieGenBench 30s | Dynamic degree: 34.69 -> 52.56 | Improved |
| Add motion-disentangling regularization on top of routing (ablation) | Self-Forcing + Frame Sink · MovieGenBench 30s | Dynamic degree: 52.56 -> 77.19 | Improved |
| Replace universal frame sink with Disen-Forcing | Self-Forcing · MovieGenBench 30s | Dynamic degree: 35.66 -> 73.12 | Improved |
| Use frame sink on Causal-Forcing | Causal-Forcing · MovieGenBench 30s | Dynamic degree: 68.28 -> 17.75 | Hurt |
| Use frame sink at the 5s training horizon | Self-Forcing · VBench 5s | Dynamic degree: 65.00 -> 37.00 | Hurt |
| Replace identified anchor heads with random heads (ablation) | Disen-Forcing · MovieGenBench 30s | Dynamic degree: 73.12 -> 51.16 (layer-matched), 34.69 (global) | Hurt |
| Lower anchor threshold to p=70 (more anchors) (ablation) | Disen-Forcing · MovieGenBench 30s | Dynamic degree: 77.19 -> 48.22 | Hurt |
| Raise anchor threshold to p=90 (fewer anchors) (ablation) | Disen-Forcing · MovieGenBench 30s | imaging quality drift: 4.39 -> 4.77 | Hurt |
| Increase regularization weight from 0.002 to 0.005 (ablation) | Disen-Forcing · MovieGenBench 30s | Dynamic degree and drift: 73.12 -> 75.97 dynamic; drift 4.03 -> 4.47 | Mixed |
| Remove VAE re-encoding of counterfactual context (ablation) | Disen-Forcing · MovieGenBench 30s | imaging quality drift: 4.03 -> 4.85 | Hurt |
| Mask top-20% anchor heads at inference (ablation) | Self-Forcing · MovieGenBench | consistency | Hurt |
| Apply regularization to four layers instead of layer 26 (ablation) | Disen-Forcing · MovieGenBench 30s | Dynamic degree: 73.12 -> 78.38; drift 4.03 -> 4.83 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Causal-Forcing — Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation (ICML 2026) · better on 3 of 3 shared results
- Deep-Forcing — Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression (ICML 2026) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Builds on: frame sink attention sinks
Closest papers by meaning
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.91)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.91)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.90)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (CVPR 2025 · similarity 0.90)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.90)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · 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.
