Glance Forcing cuts video distillation trajectory cost from thousands of GPU-hours to one

Zhuobai Dong · hf · 2026-10-15

Researchers from Wuhan University, NUS and others propose Glance Forcing, drastically cutting the cost of distilling bidirectional diffusion models into causal few-step students — the backbone of real-time video synthesis and video world models.

Problem: existing pipelines need offline PF-ODE trajectories costing tens of thousands of GPU-hours and terabytes of storage. Pushing supervision to a single sample causes severe quality collapse: the teacher trajectory's high-noise regime (semantic layout, motion) and low-noise regime (fine texture) have fundamentally different curvature, so one monolithic student suffers representation interference.

Method: decouple the trajectory into slow and fast complementary experts with stage-specific objectives, keeping the teacher frozen. This reveals a general mechanism: PF-ODE distillation mainly straightens few-step trajectory geometry, while distribution matching independently governs distributional quality.

Results: with a single PF-ODE trajectory it beats the fully-trained SOTA baseline, raising VBench Total from 84.04 to 84.22 while cutting trajectory collection from thousands of GPU-hours to about one A100-GPU-hour, with larger gains in dynamic degree, VisionReward and instruction-following.

Original post →

More from Multimodal

Multimodal channel →