Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance

Glance Forcing. Making Bidirectional Diffusion Autoregressive at a Glance. By Zhuobai Dong and five others, from Wuhan University, the University of Electronic Science and Technology of China, and four other institutions. Published on the fourteenth of October. Its code repository is worked out: about one hundred and twenty source files, with a readme and a licence.

Autoregressive generation is what real-time, causal video synthesis rests on, and it is the backbone of video world models. The dominant way to get such a model today is to distill a bidirectional diffusion model into a causal student that needs only a few steps. The conversion is the expensive part. Existing pipelines require offline Probability Flow ODE trajectories, PF-ODE for short, and collecting them costs tens of thousands of GPU-hours and terabytes of storage.

The authors argue the conversion can be made dramatically cheaper, and they push the idea to its limit. They cut the trajectory supervision down to a single sample, the strictest setting possible. That causes severe quality collapse. The cause, they find, lies in the geometry of the teacher's trajectory. The high-noise regime governs semantic layout and motion. The low-noise regime governs fine texture. The two show fundamentally different curvature and optimization behaviour, so one monolithic student cannot fit both together, and the attempt produces severe representation interference.

Glance Forcing is a lightweight distillation framework built on that diagnosis. Instead of one student, it decouples the denoising trajectory into two complementary experts, a slow one and a fast one. Each is guided by its own objective, tailored to its stage of the trajectory. Throughout, the pretrained teacher stays entirely frozen.

Splitting the work this way also exposes something more general about few-step distillation. The authors find that PF-ODE distillation mainly straightens the geometry of the few-step trajectory, while the stage that follows, distribution matching, independently governs how good the resulting distribution is. In other words, the two stages do different jobs. One is about the shape of the path the student takes from noise to video, and the other is about whether what arrives at the end looks like the real distribution. The authors present this as a general mechanism underlying few-step distillation, not only a property of their own framework.

Using just one PF-ODE trajectory, Glance Forcing already surpasses the fully trained state-of-the-art baseline. The VBench Total score rises from eighty-four point zero four to eighty-four point two two. Meanwhile the cost of collecting trajectories falls from thousands of GPU-hours to about one A100-GPU-hour. The quality number moves slightly upward. The cost number moves by orders of magnitude, and that contrast is the core of the result: matching and edging past the baseline is achieved with a sliver of the data it relied on.

The gains are markedly larger on dynamic degree, on VisionReward, and on instruction-following, and human raters consistently preferred Glance Forcing over strong baselines.

The move that makes this work is refusing to ask one student to do two incompatible things. Once the high-noise and low-noise regimes each get their own expert and their own objective, the single trajectory that collapsed a monolithic student becomes enough, and the expensive offline collection step nearly disappears.