Autoregressive video generation typically relies on distilling a bidirectional diffusion model into a causal few-step student. The standard approach depends on offline Probability Flow ODE (PF-ODE) trajectories, which are prohibitively expensive to collect—costing tens of thousands of GPU-hours and terabytes of storage. The authors of this paper set out to push that cost down, eventually reducing trajectory supervision to a single sample, the strictest possible setting.
That extreme reduction causes severe quality collapse. The paper traces the failure to the teacher's trajectory geometry: the high-noise regime, which governs semantic layout and motion, and the low-noise regime, which governs fine texture, exhibit fundamentally different curvature and optimization behavior. A single monolithic student cannot jointly fit both, leading to representation interference. Glance Forcing addresses this by splitting the denoising trajectory into two complementary experts—one slow, one fast—each guided by a stage-specific objective while the pretrained teacher remains frozen.
The method also 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 surpasses a fully-trained state-of-the-art baseline, raising the VBench Total score from 84.04 to 84.22. The gains are markedly larger in dynamic degree, VisionReward, and instruction-following, and human raters consistently preferred its outputs over strong baselines.