去噪扩散概率模型

Denoising Diffusion Probabilistic Models

乔纳森·何 Jonathan Ho · UC Berkeley · 2020-06-19 · arXiv:2006.11239 ↗ · 被引 32563

摘要 · Abstract

我们利用扩散概率模型（一类受非平衡热力学启发的潜变量模型）展示了高质量的图像合成结果。我们的最佳结果是通过在加权变分界上训练获得的，该变分界是根据扩散概率模型与使用朗之万动力学的去噪分数匹配之间的新联系设计的。我们的模型自然地允许一种渐进式有损解压方案，可解释为自回归解码的推广。在无条件 CIFAR10 数据集上，我们获得了 9.46 的 Inception 分数和 3.17 的最先进 FID 分数。在 256x256 LSUN 上，我们获得了与 ProgressiveGAN 相当的样本质量。我们的实现可在 https://github.com/hojonathanho/diffusion 获取。

We present high quality image synthesis results using diffusion probabilistic models, a class of latent variable models inspired by considerations from nonequilibrium thermodynamics. Our best results are obtained by training on a weighted variational bound designed according to a novel connection between diffusion probabilistic models and denoising score matching with Langevin dynamics, and our models naturally admit a progressive lossy decompression scheme that can be interpreted as a generalization of autoregressive decoding. On the unconditional CIFAR10 dataset, we obtain an Inception score of 9.46 and a state-of-the-art FID score of 3.17. On 256x256 LSUN, we obtain sample quality similar to ProgressiveGAN. Our implementation is available at https://github.com/hojonathanho/diffusion

核心贡献 · Key contributions

建立了扩散模型与基于朗之万动力学的去噪分数匹配之间的联系。
Established connection between diffusion models and denoising score matching with Langevin dynamics.
提出了反向过程的 epsilon 预测参数化，提升了样本质量。
Proposed epsilon-prediction parameterization for reverse process, improving sample quality.
在无条件 CIFAR10 上实现了 3.17 的 FID 分数，达到最先进水平。
Achieved state-of-the-art FID score of 3.17 on unconditional CIFAR10.
展示了渐进式有损解压缩方案，推广了自回归解码。
Demonstrated progressive lossy decompression scheme generalizing autoregressive decoding.
表明扩散模型作为有损压缩器，对不可察觉细节具有归纳偏置。
Showed diffusion models as lossy compressors with inductive bias for imperceptible details.
提供了开源实现以促进可重复性。
Provided open-source implementation for reproducibility.

局限 · Limitations

对数似然与其他基于似然的模型相比不具备竞争力。
Log-likelihoods not competitive with other likelihood-based models.
样本质量仅限于图像数据；未探索其他模态。
Sample quality limited to image data; other modalities not explored.
在简化目标上训练产生更好的样本但码长更差。
Training on simplified objective yields better samples but worse codelengths.
反向过程方差固定为常数，限制了表达能力。
Reverse process variances fixed as constants, limiting expressiveness.
可能被恶意用于生成虚假图像或强化偏见。
Potential for malicious use in generating fake images or reinforcing biases.

论文章节 · Sections（共 17）

摘要 Abstract
1 引言 1 Introduction
2 背景 2 Background
3 扩散模型与去噪自编码器 3 Diffusion models and denoising autoencoders
3.1 前向过程与 LT 3.1 Forward process and LTsubscript𝐿𝑇L_{T}
3.2 反向过程与 L1:T-1 3.2 Reverse process and L1:T−1subscript𝐿:1𝑇1L_{1:T-1}
3.3 数据缩放、反向过程解码器与 L0 3.3 Data scaling, reverse process decoder, and L0subscript𝐿0L_{0}
3.4 简化训练目标 3.4 Simplified training objective
4 实验 4 Experiments
4.1 样本质量 4.1 Sample quality
4.2 反向过程参数化与训练目标消融 4.2 Reverse process parameterization and training objective ablation
4.3 渐进式编码 4.3 Progressive coding
4.4 插值 4.4 Interpolation
5 相关工作 5 Related Work
6 结论 6 Conclusion
广泛影响 Broader Impact
致谢与资金披露 Acknowledgments and Disclosure of Funding

阅读逐段中英对照全文 →