Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance

📄 arXiv: 2610.20220v1 📥 PDF

作者: Zhuobai Dong, Junchao Yi, Hu Jian Guo, Linjie Li, Rui Zhao, Alex Jinpeng Wang

分类: cs.CV

发布日期: 2026-10-14

💡 一句话要点

提出Glance Forcing以解决双向扩散模型的高成本问题

🎯 匹配领域: 支柱二：RL算法与架构 (RL & Architecture) 支柱九：具身大模型 (Embodied Foundation Models)

关键词: 自回归生成 双向扩散模型 蒸馏训练 视频合成 高效算法

📋 核心要点

  1. 现有方法在将双向扩散模型转化为因果模型时，需耗费大量计算资源和存储，效率低下。
  2. 本文提出Glance Forcing，通过将去噪轨迹解耦为慢速和快速两个专家，优化了蒸馏过程。
  3. 实验结果显示，Glance Forcing在VBench Total评分上从84.04提升至84.22，且大幅降低了计算成本。

📝 摘要（中文）

自回归生成是实时因果视频合成的基础，然而现有的双向扩散模型转化为因果模型的过程成本极高，需耗费数万GPU小时和大量存储。本文提出Glance Forcing，一个轻量级蒸馏框架，通过将去噪轨迹解耦为慢速和快速两个专家，显著提高了效率。实验表明，使用单个PF-ODE轨迹，Glance Forcing在多个指标上超越了现有最优基线，同时将轨迹收集成本降低至约一小时A100-GPU的水平。

🔬 方法详解

问题定义：现有的双向扩散模型在转化为因果自回归模型时，依赖于高成本的PF-ODE轨迹，导致效率低下和资源浪费。

核心思路：本文通过引入Glance Forcing框架，将去噪轨迹分为慢速和快速两个专家，分别针对不同的优化目标进行训练，从而提高蒸馏效率。

技术框架：该框架包括两个主要模块：慢速专家和快速专家。慢速专家负责处理高噪声的语义布局和运动，而快速专家则专注于低噪声的细节纹理。整个过程保持预训练教师模型不变。

关键创新：Glance Forcing的创新在于将去噪轨迹解耦为两个互补的专家，解决了单一模型无法同时适应不同轨迹几何形状的问题，显著提升了生成质量。

关键设计：在损失函数设计上，慢速和快速专家各自采用了特定的阶段性目标，确保了在不同噪声水平下的优化效果，同时使用单个PF-ODE轨迹进行训练，极大降低了计算成本。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果显示，Glance Forcing在VBench Total评分上从84.04提升至84.22，超越了现有最优基线。同时，轨迹收集成本从数千GPU小时降低至约一小时A100-GPU，动态度、VisionReward和指令跟随等指标也有显著提升。

🎯 应用场景

该研究的潜在应用领域包括实时视频合成、虚拟现实和增强现实等场景，能够为这些领域提供更高效的生成模型，提升用户体验。未来，Glance Forcing可能在多模态生成和智能内容创作中发挥重要作用。

📄 摘要（原文）

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.