以弱到强分段引导提升扩散模型泛化能力
Improving Diffusion Generalization with Weak-to-Strong Segmented Guidance
把 CFG 与 AutoGuidance 统一到弱到强原则下，给出何时该用哪种引导的判据，并提出 SEG 混合引导；做扩散采样或引导调参时可直接替换 CFG，也可把该原则写进训练目标以提升无引导模型的泛化。
原文依据 we propose a hybrid instantiation called SEG under the principle, taking the benefits of both
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论扩散模型通过迭代精炼过程生成合成图像。然而，无模拟目标与迭代过程之间的错位，往往会在采样轨迹上造成累积的梯度误差，从而导致结果不理想并无法泛化。分类器无引导（CFG）与自动引导（AG）等引导技术通过在主导信号与次优信号之间进行外推来缓解这一问题，以获得更强的泛化能力。尽管这些方法在经验上取得了成功，但主流引导方法的有效作用区间仍未被充分探索，导致在给定前提条件下难以明确该选择哪种引导方法。在本工作中，我们首先进行合成对比实验，从弱到强原则的视角分离并展示以 CFG 和 AG 为代表的引导方法的有效作用区间。在此基础上，我们在该原则下提出一种名为 SEG 的混合实例化方法，兼取两者之长。进一步地，我们证明弱到强原则与 SEG 可以被迁移到训练目标中，从而提升无引导扩散模型的泛化能力。我们通过全面的实验验证了所提方法。在推理阶段，在 SD3 与 SD3.5 上的评估证实 SEG 优于现有的无训练引导变体。在 Transformer 架构上的训练阶段实验表明，该迁移在条件与无条件设定下均有效，并带来性能提升。代码见 https://github.com/Westlake-AGI-Lab/SGG。
英文原文摘要 1577 字符
Diffusion models generate synthetic images through an iterative refinement process. However, the misalignment between the simulation-free objective and the iterative process often causes accumulated gradient error along the sampling trajectory, which leads to unsatisfactory results and a failure to generalize. Guidance techniques like Classifier Free Guidance (CFG) and AutoGuidance (AG) alleviate this by extrapolating between the main and inferior signal for stronger generalization. Despite empirical success, the effective operational regimes of prevalent guidance methods are still under-explored, leading to ambiguity when selecting the appropriate guidance method given a precondition. In this work, we first conduct synthetic comparisons to isolate and demonstrate the effective regime of guidance methods represented by CFG and AG from the perspective of weak-to-strong principle. Based on this, we propose a hybrid instantiation called SEG under the principle, taking the benefits of both. Furthermore, we demonstrate that the W2S principle along with SEG can be migrated into the training objective, improving the generalization ability of unguided diffusion models. We validate our approach with comprehensive experiments. At inference time, evaluations on SD3 and SD3.5 confirm that SEG outperforms existing training-free guidance variants. Training-time experiments on transformer architectures demonstrate the effective migration and performance gains in both conditional and unconditional settings. Code is available at https://github.com/Westlake-AGI-Lab/SGG