The primary axes of interest in image-generating diffusion models are image quality, the amount of variation in the results, and how well the results align with a given condition, e.g., a class label or a text prompt. The popular classifier-free guidance approach uses an unconditional model to guide a conditional model, leading to simultaneously better prompt alignment and higher-quality images at the cost of reduced variation. These effects seem inherently entangled, and thus hard to control. We make the surprising observation that it is possible to obtain disentangled control over image quality without compromising the amount of variation by guiding generation using a smaller, less-trained version of the model itself rather than an unconditional model. This leads to significant improvements in ImageNet generation, setting record FIDs of 1.01 for 64x64 and 1.25 for 512x512, using publicly available networks. Furthermore, the method is also applicable to unconditional diffusion models, drastically improving their quality.
核心贡献 · Key contributions
提出自引导方法，使用模型自身更小、训练更少的版本来引导生成，在不减少多样性的情况下实现对图像质量的解耦控制。 Proposes autoguidance, using a smaller, less-trained version of the model itself to guide generation, achieving disentangled control over image quality without reducing variation.
在 ImageNet 上创下新的最先进 FID 记录：64x64 为 1.01，512x512 为 1.25。 Sets new state-of-the-art FID records on ImageNet: 1.01 for 64x64 and 1.25 for 512x512.
发现无分类器引导通过条件与无条件去噪器之间的质量差异，隐式截断低概率区域，从而提升图像质量。 Identifies that classifier-free guidance improves image quality by implicitly truncating low-probability regions due to quality differences between conditional and unconditional denoisers.
证明自引导可应用于无条件扩散模型，大幅提升其质量（例如 ImageNet-512 上 FID 从 11.67 降至 3.86）。 Demonstrates that autoguidance is applicable to unconditional diffusion models, drastically improving their quality (e.g., FID from 11.67 to 3.86 on ImageNet-512).
验证引导模型必须遭受兼容的退化（如降低容量和训练）才能有效引导。 Validates that the guiding model must suffer from compatible degradations (e.g., reduced capacity and training) for effective guidance.
表明自引导比无分类器引导更好地保留图像风格和复杂性，并可与之结合用于艺术控制。 Shows autoguidance preserves image style and complexity better than classifier-free guidance, and can be combined with it for artistic control.
局限 · Limitations
需要同一模型的更小、训练更少的版本，现有大规模生成器可能无法提供。 Requires a smaller, less-trained version of the same model, which may not be available for existing large-scale generators.
有效性取决于主模型与引导模型之间的兼容退化；不匹配的退化不会带来改进。 Effectiveness depends on compatible degradations between main and guiding models; mismatched degradations yield no improvement.
未提供有益自引导条件的正式证明，依赖经验观察。 Does not provide formal proof of conditions for beneficial autoguidance; relies on empirical observations.
训练快照之间的潜在分布偏移（例如由于数据变化）可能违反假设。 Potential distribution shifts between training snapshots (e.g., due to data changes) may violate assumptions.
最佳引导模型选择（如容量、训练步数）需要网格搜索，且可能因数据集而异。 Optimal guiding model selection (e.g., capacity, training steps) requires grid search and may vary across datasets.
论文章节 · Sections（共 10）
摘要Abstract
1 引言1 Introduction
2 背景2 Background
3 为什么 CFG 能提升图像质量？3 Why does CFG improve image quality?