We show that diffusion models can achieve image sample quality superior to the current state-of-the-art generative models. We achieve this on unconditional image synthesis by finding a better architecture through a series of ablations. For conditional image synthesis, we further improve sample quality with classifier guidance: a simple, compute-efficient method for trading off diversity for fidelity using gradients from a classifier. We achieve an FID of 2.97 on ImageNet 128$\times$128, 4.59 on ImageNet 256$\times$256, and 7.72 on ImageNet 512$\times$512, and we match BigGAN-deep even with as few as 25 forward passes per sample, all while maintaining better coverage of the distribution. Finally, we find that classifier guidance combines well with upsampling diffusion models, further improving FID to 3.94 on ImageNet 256$\times$256 and 3.85 on ImageNet 512$\times$512. We release our code at https://github.com/openai/guided-diffusion
核心贡献 · Key contributions
证明扩散模型在无条件与条件图像合成中样本质量超越 GANs。 Demonstrates diffusion models surpass GANs in image sample quality on unconditional and conditional synthesis.
提出改进的 UNet 架构，包含多分辨率注意力、BigGAN 模块和自适应组归一化。 Proposes improved UNet architecture with multi-resolution attention, BigGAN blocks, and adaptive group normalization.
引入分类器引导，通过分类器梯度在多样性与保真度间权衡。 Introduces classifier guidance for trading off diversity for fidelity using classifier gradients.
在 ImageNet 多个分辨率上以少量采样步骤取得最优 FID 分数。 Achieves state-of-the-art FID scores on ImageNet at multiple resolutions with few sampling steps.
展示分类器引导与上采样扩散模型结合可进一步改善 FID。 Shows classifier guidance combines with upsampling diffusion models for further FID improvements.
提供引导扩散模型的开源代码。 Provides open-source code for guided diffusion models.
局限 · Limitations
因多次前向传播，采样速度慢于 GANs。 Sampling is slower than GANs due to multiple forward passes.
分类器引导需要标注数据集，缺乏针对无标注数据的有效策略。 Classifier guidance requires labeled datasets; no effective strategy for unlabeled data.
FID 等指标并非人类判断的完美代理。 Metrics like FID are imperfect proxies for human judgment.
需要调整引导尺度以优化多样性与保真度的权衡。 Guidance scale tuning is needed for optimal trade-off between diversity and fidelity.
架构改进可能无法不经调优直接迁移至其他领域。 Architecture improvements may not transfer to other domains without tuning.
论文章节 · Sections（共 18）
摘要Abstract
1 引言1 Introduction
2 背景2 Background
2.1 改进2.1 Improvements
2.2 样本质量指标2.2 Sample Quality Metrics
3 架构改进3 Architecture Improvements
3.1 自适应组归一化3.1 Adaptive Group Normalization
4 分类器引导4 Classifier Guidance
4.1 条件反向加噪过程4.1 Conditional Reverse Noising Process