DDPM:去噪扩散概率模型Denoising Diffusion Probabilistic Models
问题:GAN 能出高质量样本但训练不稳定,VAE 有似然但样本偏模糊;能不能用「逐步加噪、再逐步去噪」的扩散过程,同时拿到质量和可训练性?
证据位置:前向/反向过程在 §2–§3,训练目标(简化噪声预测损失)在 §3,采样在 Algorithm 2;实验在 §4,CIFAR10 的 IS/FID 在 Table 1,LSUN 256×256 在 §4.4 左右;结论在 §6,影响讨论在 Broader Impact。
作者自陈局限:论文没有独立 Limitations 章节。摘要与 §1 明确说「样本质量好,但似然不如其他似然模型有竞争力」,并解释大部分无损码长花在不可感知的细节上(§4.3);Broader Impact 主动写了两类风险——可被用于伪造公众人物图像与视频,以及训练数据偏见会被生成模型放大。
它没主张的:没有说扩散模型在似然上优于 ARM/VAE;没有说采样已经够快——扩散需要多步迭代,这一点是后续工作(DDIM、蒸馏等)才改善;也没有说 FID 3.17 可以跨数据集比较——FID 依赖特征提取器与样本量,论文只报了自己的设置。Problem: GANs produce sharp samples but train unstably, while VAEs have likelihoods but blurrier samples; can a process that gradually adds noise and then denoises it deliver both quality and trainability?
Where the evidence sits: forward and reverse processes in Sections 2-3, the simplified noise-prediction objective in Section 3, sampling in Algorithm 2; experiments in Section 4, with CIFAR10 IS/FID in Table 1 and 256x256 LSUN around 4.4; conclusion in Section 6 and societal impact in the Broader Impact section.
Stated limits: there is no separate Limitations section. The abstract and Section 1 state plainly that sample quality is high but likelihoods are not competitive with other likelihood-based models, explaining that most lossless code length goes to imperceptible detail (Section 4.3); the Broader Impact section volunteers two risks — use for fake images and videos of public figures, and amplification of dataset biases.
What it does not claim: it does not claim diffusion beats autoregressive models or VAEs on likelihood; it does not claim sampling is fast — diffusion needs many iterative steps, which later work (DDIM, distillation) addressed; and it does not claim FID 3.17 is comparable across datasets, since FID depends on the feature extractor and sample size.
关键参数Key specs
- 证据位置 / evidence
- §2–3 扩散过程;§4 实验(Table 1);§4.3 码长分析;§6 结论;Broader Impact
- 规模 / scale
- CIFAR10 无条件生成;256×256 LSUN
- 关键数字 / key numbers
- CIFAR10 IS 9.46、FID 3.17;LSUN 样本质量接近 ProgressiveGAN
- 自陈局限 / stated limits
- 似然不具竞争力;伪造图像/视频与偏见放大风险;无独立 Limitations 章节
- 复现状态 / reproduction
- 未复现
官方原文怎么说What the source says
On the unconditional CIFAR10 dataset, we obtain an Inception score of 9.46 and a state-of-the-art FID score of 3.17.
our models do not have competitive log likelihoods compared to other likelihood-based models
Sample generation techniques can be employed to produce fake images and videos of high profile figures for political purposes.
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-17
- 许可 / 类型Licence / kind
- arxiv-preprint
- 许可说明Licence evidence
- https://arxiv.org/abs/2006.11239
- 类目Category
- 架构Architecture
论文里的数字怎么读How to read the numbers
先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards