Glance forcing improves fast video synthesis with less computing cost
Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
Computer Vision and Pattern Recognition
Summary
Making videos using AI models often requires a lot of computing power and time, especially when these models have to predict things in order. The authors show that the usual ways to teach smaller, faster models from bigger ones are very costly and slow. They find that trying to simplify this process too much causes problems because different parts of video generation behave differently. To fix this, they introduce a method called Glance Forcing that splits the problem into two easier parts, making training much faster without losing quality. This approach performs better than previous methods while needing far fewer resources.
What this means in practice
- •For video synthesis engineers: Cut the computing time and resources needed to train real-time video generation models by using Glance Forcing for efficient distillation.
- •For machine learning platform teams: Reduce storage and GPU usage significantly when preparing lightweight video generation models for deployment using the proposed distillation approach.
Authors
Zhuobai Dong, Junchao Yi, Hu Jian Guo, Linjie Li, Rui Zhao, Alex Jinpeng Wang
Abstract
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.