NeurIPS 2026

Disen-Forcing

Disentangling Semantic Anchoring from Motion for Autoregressive Video Diffusion

Zhengyang Yu^1,★Akio Hayakawa²Masato Ishii²Qingtao Yu¹ Xuesong Li^1,3Takashi Shibuya²Yuki Mitsufuji^2,4Jing Zhang¹

¹Australian National University · ²Sony AI · ³CSIRO · ⁴Sony Group Corporation

^★Zhengyang Yu is looking for research internship and full-time opportunities: zhengyang.yu@anu.edu.au

Paper Code · coming soon BibTeX

Training dynamics of Dynamic Degree and Imaging Quality, and a 30-second qualitative comparison of Self-Forcing, Frame Sink and Disen-Forcing

TL;DR

Diagnosis

Initial-frame tokens are not an LLM-style attention sink in AR-VDMs: long-range consistency is anchored by a sparse subset of attention heads, while most heads capture local temporal dynamics. Frame sink exposes every head to the initial frames, which makes distillation prone to collapsing motion.

Method

Disen-Forcing is a distillation-time plug-in that disentangles semantic anchoring from motion in AR-VDMs via disentangled context routing and motion-disentangling regularization.

Results

On long-horizon video generation, Disen-Forcing improves both the long-range consistency and the motion dynamics of base AR-VDMs, a trade-off that sink-based methods fail to balance.

Diagnosis

Initial-frame tokens are not a universal attention sink

The vast majority of heads attend predominantly to the current frame, not the initial frames.

Per-head attention to the initial and current frames across six AR-VDMs — Per-head attention to the initial and current frames across AR-VDMs, averaged over 50 MovieGenBench prompts and denoising timesteps. Most heads attend predominantly to the current frame, while a sparse subset concentrates attention on the initial frame.

Initial frames sustain consistency through sparse heads

Most heads are Working Heads with a local temporal pattern; a sparse minority of Anchor Heads attend strongly to the initial frame across AR steps.

Attention distributions of representative Self-Forcing heads across seven rollout steps — Representative Self-Forcing heads across seven rollout steps, at ODE initialization (left) and after DMD (right). Working Heads exhibit local temporal patterns, whereas Anchor Heads attend persistently to the initial frame.

Masking anchor heads versus random heads: consistency metrics and frames over 5 seconds — Masking the top 20% Anchor Heads leads to a sharp drop in subject/background consistency and a clear rise in imaging-quality drift, together with visible appearance drift **even within the 5-second training horizon**. Masking the same number of Working Heads at random yields metrics on par with the unmasked baseline and no observable drift.

Universal frame sink causes motion collapse

In current sink-based AR-VDMs, however, the initial frames are exposed to every head. Combined with the mode-seeking DMD objective, this offers the student an easy shortcut toward high-fidelity static modes, leading it to over-condition on the initial frames at the cost of motion.

Method

Given an ODE-initialization model, we first perform a one-time identification of anchor heads, then distill with two complementary components.

Disentangled context routing

Motion-disentangling regularization

Routing alone cannot fully close the leakage from Anchor Heads into motion construction: it constrains only which context each head reads, not which aspect of the generated video each head shapes. We therefore require local motion construction to remain invariant under intervention on Anchor Heads' access to initial frames. Using the cross-frame self-attention map as a motion proxy, we run a counterfactual pass that replaces the anchor heads' routed context with the sliding window, and penalize the KL divergence between the two proxies in layers that contain no anchor heads.

Results

Self-Forcing suffers from temporal drift once the inference rollout exceeds the training horizon; Frame Sink avoids temporal drift but yields low-dynamic, conservative motion. Ours (Disen-Forcing) is both drift-free and temporally dynamic.

30-second generation

1-minute generation

Quantitative comparison

Base models, Frame Sink and Disen-Forcing at 30 seconds and 1 minute
Model	30 s			1 min
Model	Imaging Quality ↑	Dynamic Degree ↑	Quality Drift ↓	Imaging Quality ↑	Dynamic Degree ↑	Quality Drift ↓
Self-Forcing	67.87	40.44	7.23	65.68	34.70	12.73
+ Frame Sink	69.94	35.66	3.40	70.34	35.83	3.10
+ Disen-Forcing	70.05	73.12	4.03	70.01	73.95	4.04
Causal-Forcing	65.75	68.28	9.38	62.57	64.91	12.96
+ Frame Sink	71.45	17.75	3.25	71.51	16.06	3.66
+ Disen-Forcing	69.77	69.72	4.83	68.96	69.64	5.10

BibTeX

@article{yu2026disenforcing,
  title={Disen-Forcing: Disentangling Semantic Anchoring from Motion for Autoregressive Video Diffusion},
  author={Yu, Zhengyang and Hayakawa, Akio and Ishii, Masato and Yu, Qingtao and Li, Xuesong and Shibuya, Takashi and Mitsufuji, Yuki and Zhang, Jing},
  journal={arXiv preprint arXiv:2611.02822},
  year={2026}
}