Video World Models · 2026-10-15
Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
Zhuobai Dong, Junchao Yi, Hu Jian Guo, Linjie Li, Rui Zhao, Alex Jinpeng Wang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors propose Glance Forcing, a lightweight distillation method that makes bidirectional diffusion video models causal and few-step with much less PF-ODE trajectory supervision. When supervision is reduced to one PF-ODE trajectory, prior methods collapse due to conflicting high-noise vs low-noise optimization regimes. Glance Forcing decouples the denoising trajectory into two phase experts (Slow/Fast) using dual-LoRA with phase-specific objectives, while freezing the pretrained teacher.
Why it matters
Autoregressive video generation is the backbone of video world models, but existing bidirectional-to-causal distillation pipelines require expensive offline PF-ODE trajectories. The paper shows extreme sparsity (one trajectory) causes severe quality collapse, and attributes it to representation interference from different trajectory geometry/optimization behavior across denoising phases. Glance Forcing targets this with a phase-aware dual-LoRA distillation design and demonstrates improved VBench quality with greatly reduced trajectory-collection cost.
Method
- Propose Glance Forcing, a phase-aware distillation framework that decouples the denoising trajectory into two complementary Slow/fast experts using dual-LoRA, while keeping the pretrained teacher entirely frozen.
- Instead of costly offline PF-ODE trajectory collection, the method learns from a single sample trajectory; it uses stage-specific objectives (Temporal Frequency Spectrum Loss for Slow; Spatial High-Pass Loss for Fast).
- The authors characterize why one-sample supervision fails for a single monolithic student: high-noise (semantic layout/motion) and low-noise (fine texture) regimes have different curvature and optimization behavior, causing representation interference.
Abstract (from arXiv)
Autoregressive generation underlies real-time, causal video synthesis and is the backbone of video world models. Today's dominant approach distills a bidirectional diffusion model into a causal few-step student, but existing pipelines make this conversion prohibitively expensive, requiring offline Probability Flow ODE} (PF-ODE) trajectories that cost tens of thousands of GPU-hours and terabytes of storage. We show it can be made dramatically more efficient. Pushing this efficiency to its extreme, we reduce trajectory supervision to a single sample, the strictest setting possible, and find that it causes severe quality collapse. We trace this to the teacher's trajectory geometry: the high-noise regime, governing semantic layout and motion, and the low-noise regime, governing fine texture exhibit fundamentally different curvature and optimization behavior, so a single monolithic student cannot jointly fit both, causing severe representation interference. Motivated by this, we propose Glance Forcing, a lightweight distillation framework that decouples the denoising trajectory into two complementary slow and fast experts, each guided by a tailored, stage-specific objective, while keeping the pretrained teacher entirely frozen. This decoupling further reveals a general mechanism underlying few-step distillation: PF-ODE distillation mainly straightens the few-step trajectory geometry, while the subsequent distribution-matching stage independently governs distributional quality. Using only one PF-ODE trajectory, Glance Forcing already surpasses the fully-trained state-of-the-art baseline, raising the VBench Total score from 84.04 to 84.22 while cutting trajectory-collection cost from thousands of GPU-hours to about one A100-GPU-hour, with markedly larger gains in dynamic degree, VisionReward, and instruction-following, and consistent preference from human raters over strong baselines.
Related papers
- LongTake: Learning to Sustain Dynamics in Long-Horizon Video Generation
- Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
- LongForcing: Scaling Temporal Manifold Matching for Long-Horizon World Rollout
- Rethinking Video Autoencoders for Interactive World Models
- MoDDiT: Simple Token Routing in Diffusion Transformers
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-14 · Indexed by Pub-AI: 2026-10-15 UTC
Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
Preprint posted: 2026-10-14 · Original record ↗
- arxiv
- 2610.20220
Explicit source versions (1)
- v1 · 2026-10-14
- Preprint available
Source date: 2026-10-14 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Autoregressive generation underlies real-time, causal video synthesis and is the backbone of video world models. Today's dominant approach distills a bidirectional diffusion model into a causal few-step student, but existing pipelines make this conversion prohibitively expensive, requiring offline Probability Flow ODE} (PF-ODE) trajectories that cost tens of thousands of GPU-hours and terabytes of storage. We show it can be made dramatically more efficient. Pushing this efficiency to its extreme, we reduce trajectory supervision to a single sample, the strictest setting possible, and find that it causes severe quality collapse. We trace this to the teacher's trajectory geometry: the high-noise regime, governing semantic layout and motion, and the low-noise regime, governing fine texture exhibit fundamentally different curvature and optimization behavior, so a single monolithic student cannot jointly fit both, causing severe representation interference. Motivated by this, we propose Glance Forcing, a lightweight distillation framework that decouples the denoising trajectory into two complementary slow and fast experts, each guided by a tailored, stage-specific objective, while keeping the pretrained teacher entirely frozen. This decoupling further reveals a general mechanism underlying few-step distillation: PF-ODE distillation mainly straightens the few-step trajectory geometry, while the subsequent distribution-matching stage independently governs distributional quality. Using only one PF-ODE trajectory, Glance Forcing already surpasses the fully-trained state-of-the-art baseline, raising the VBench Total score from 84.04 to 84.22 while cutting trajectory-collection cost from thousands of GPU-hours to about one A100-GPU-hour, with markedly larger gains in dynamic degree, VisionReward, and instruction-following, and consistent preference from human raters over strong baselines.