Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
Glance Forcing distills causal video diffusion with one PF-ODE trajectory and dual slow-fast experts.
Glance Forcing is a lightweight distillation method that turns a frozen bidirectional video diffusion teacher into a causal few-step student. It splits the denoising trajectory into slow and fast experts so one Probability Flow ODE sample can supervise both semantic motion and fine texture. Using a single trajectory, it raises VBench Total from 84.04 to 84.22 and cuts trajectory-collection cost from thousands of GPU-hours to about one A100-hour, with gains in dynamic degree, VisionReward, and instruction following.
- One PF-ODE sample collapses quality because high- and low-noise regimes conflict.
- Glance Forcing uses frozen-teacher slow and fast experts with stage-specific objectives.
- VBench Total rises from 84.04 to 84.22 versus a fully trained baseline.
- Trajectory collection drops from thousands of GPU-hours to about one A100-hour.
From the original article
1 min read at huggingface.co
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.
By Zhuobai Dong, Junchao Yi, Hu Jian Guo · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.