Classifier guidance is a recently introduced method to trade off mode coverage and sample fidelity in conditional diffusion models post training, in the same spirit as low temperature sampling or truncation in other types of generative models. Classifier guidance combines the score estimate of a diffusion model with the gradient of an image classifier and thereby requires training an image classifier separate from the diffusion model. It also raises the question of whether guidance can be performed without a classifier. We show that guidance can be indeed performed by a pure generative model without such a classifier: in what we call classifier-free guidance, we jointly train a conditional and an unconditional diffusion model, and we combine the resulting conditional and unconditional score estimates to attain a trade-off between sample quality and diversity similar to that obtained using classifier guidance.
核心贡献 · Key contributions
提出无分类器引导，一种无需分类器即可在扩散模型中权衡样本质量和多样性的方法。 Proposes classifier-free guidance, a method to trade off sample quality and diversity in diffusion models without a classifier.
通过随机丢弃条件信息，使用单一神经网络联合训练条件和无条件扩散模型。 Jointly trains conditional and unconditional diffusion models using a single neural network by randomly dropping conditioning.
证明无分类器引导在 ImageNet 上实现了与分类器引导相当的 FID/IS 权衡。 Demonstrates that classifier-free guidance achieves FID/IS tradeoffs comparable to classifier guidance on ImageNet.
表明纯生成模型可以在没有对抗性分类器梯度的情况下提升基于分类器的指标。 Shows that pure generative models can boost classifier-based metrics without adversarial classifier gradients.
提供直观解释：引导降低无条件似然同时增加条件似然。 Provides an intuitive explanation: guidance decreases unconditional likelihood while increasing conditional likelihood.
在 128x128 ImageNet 上取得最先进 FID，优于分类器引导的 ADM-G 和 BigGAN-deep。 Achieves state-of-the-art FID on 128x128 ImageNet, outperforming classifier-guided ADM-G and BigGAN-deep.
局限 · Limitations
每个采样步骤需要两次前向传播，与分类器引导相比降低了采样速度。 Requires two forward passes per sampling step, reducing sampling speed compared to classifier guidance.
依赖于训练无条件模型，对于高维条件可能效率低下。 Relies on training an unconditional model, which may be inefficient for high-dimensional conditioning.
样本多样性降低可能对部署模型产生负面影响，尤其是对代表性不足的数据。 Decreased sample diversity may have negative impacts in deployed models, especially for underrepresented data.
超参数 puncond 需要调优；最优值可能因数据集而异。 Hyperparameter puncond requires tuning; optimal value may vary across datasets.
该方法仅在 ImageNet 上验证；对其他模态的泛化性尚未探索。 Method is demonstrated only on ImageNet; generalizability to other modalities is not explored.
论文章节 · Sections（共 13）
摘要Abstract
1 引言1 Introduction
2 背景2 Background
3 引导3 Guidance
3.1 分类器引导3.1 Classifier guidance
3.2 无分类器引导3.2 Classifier-free guidance
4 实验4 Experiments
4.1 改变无分类器引导强度4.1 Varying the classifier-free guidance strength
4.2 改变无条件训练概率4.2 Varying the unconditional training probability
4.3 改变采样步数4.3 Varying the number of sampling steps