DDPM 让扩散模型成为实用生成路线

先把一张图弄成噪声，再一步步问：这一步的噪声是什么？

2020 年，Jonathan Ho、Ajay Jain 与 Pieter Abbeel 的去噪扩散概率模型（DDPM）并没有发明“扩散”这个词。它做的是把一条说得清楚的概率过程，收成当时够用的图像生成配方。前向过程固定：在论文常用的约一千步内，按预设方差日程逐步向图像添加高斯噪声，直到样本接近各向同性噪声。反向过程由网络来学——论文采用带跳跃连接的 U 型网络——在每一步估计噪声或等价地估计均值，从纯噪声迭代还原图像。关键简化是损失：不必直接优化难处理的边际似然，可对每一步噪声预测做均方误差；它与变分下界紧密相关，实现上却像普通回归。

对抗生成网络能把画面拉得很锐，却常把训练稳定性留在运气里；这篇工作想把“稳”写回损失函数。前向日程若设计不当，反向网络会学得很辛苦；论文采用的时间步与方差策略成为默认起点，后来才被余弦日程等改进替换。U 型网络的多尺度跳跃连接特别适合图像去噪任务，与扩散目标几乎一拍即合，使复现者不需要先调好一对互相对抗的网络。

在彩色小图基准上的数字给了入场券。论文报告的弗雷歇起始距离（FID）约 3.17、起始分数（Inception Score）约 9.46，与当时强对抗生成基线同场比较已具竞争力；样本质量不再是扩散方法的软肋。训练不需要判别器与生成器小心同步，模式崩塌的经典叙事随之减弱——代价是采样要跑数百到一千次网络前向，速度慢于单次前向的对抗生成。加速采样、知识蒸馏与潜空间扩散随后都在偿还这笔时间税，同时保留逐步还原的骨架。

逐步去噪让中间状态可以可视化、可以改日程、可以注入条件，这是单次生成难以提供的操作面；有了轨迹，文本条件、分类器引导与潜空间压缩才有明确挂载点。从研究生态看，该方法降低了进入高质量图像生成的训练门槛：实现一个回归噪声的网络，比调好一对互相对抗的网络更不容易在复现时彻底失败。采样慢被明确列为代价，于是加速成为正当的后续题目，而不是对原方法的羞辱。后来的多个文生图系统——DALL·E 2、Imagen、Stable Diffusion 等——共享迭代去噪骨架，尽管噪声日程、骨干网络与条件模块各不相同；那些产品是后续事件，不应倒灌成 2020 年论文的剧情预告。

生成模型的中心舞台从对抗博弈转到逐步还原，并不意味着对抗网络立刻消失，而是默认实现路径变了：先回归噪声，再迭代采样，然后才谈条件与加速。后来者常把注意力放在如何把一千步压到几十步，却容易忘记：若没有先把生成写成可逐步检查的轨迹，压缩步数将失去明确的优化目标。论文的历史贡献，是把问题从“如何让两个网络别打崩”改写成“如何更准地预测这一步的噪声”。改写之后，加速、条件与潜空间才是同一条路上的工程题目。

去噪扩散概率模型交付的是可检查的轨迹，而不只是终点图像的好看程度。一千步既是时间税，也是操作面本身。

First ruin an image with noise. Then ask, at each step: what is the noise here?

Ho, Jain, and Abbeel’s 2020 Denoising Diffusion Probabilistic Models (DDPM) did not invent the word “diffusion.” It turned a clean probabilistic process into an image-generation recipe practical enough for the moment. The forward process is fixed: over (T) steps (commonly (T=1000) in the paper), Gaussian noise is added to an image on a preset variance schedule until the sample approaches isotropic noise. The reverse process is learned: a neural network—a U-Net in the paper—estimates the noise (or an equivalent mean) at each step and iterates from pure noise back to an image. The crucial simplification is the training target. Instead of optimizing an intractable marginal likelihood directly, one can regress to the noise with mean squared error at each step—tightly linked to a variational bound, yet as stable to implement as ordinary regression.

Adversarial generators could push sharpness high and leave stability outside the recipe; this work tried to write stability back into the loss. Forward schedules, if poorly chosen, make reverse learning hard; the paper’s timestep and variance choices became defaults later improved by cosine schedules and faster samplers. U-Net skip connections fit image denoising almost naturally, so reproducers did not first have to keep a pair of adversarial networks in balance.

Numbers on CIFAR-10 bought the line a ticket into mainstream conversation. The paper reported a FID of about 3.17 and an Inception Score of about 9.46, competitive with strong GAN baselines of the period; sample quality was no longer diffusion’s soft spot. Training needed no careful discriminator–generator sync, and the classic mode-collapse story receded—at the cost of hundreds to a thousand network forwards per sample, slower than single-pass GANs. Later accelerated samplers, distillation, and latent-space diffusion all repay that compute bill while keeping the iterative-denoising skeleton.

Stepwise denoising makes intermediate states visualizable, schedules editable, and conditions injectable—an operational surface single-pass generation rarely offers. Once generation is a inspectable trajectory, text conditioning, classifier guidance, and latent compression have somewhere to attach. For the research ecology, the method lowered the floor for high-quality image generation: implementing a noise-regression net is less likely to fail completely on reproduction than balancing two adversaries. Slow sampling is listed as a cost, so acceleration becomes a legitimate research program rather than an insult to the method. Systems such as DALL·E 2, Imagen, and Stable Diffusion later share an iterative-denoising skeleton even when noise schedules, backbones, and conditioners differ—those products are later events, not a trailer written into the 2020 paper.

The center stage of generative modeling shifted from adversarial games toward stepwise recovery. GANs did not vanish overnight; the default implementation path changed: regress noise first, sample iteratively, then talk about conditioning and speed. Later work often focuses on compressing a thousand steps into dozens, and forgets that without first writing generation as a checkable trajectory, step reduction has no clear objective. The paper’s historical gift is rewriting the problem from “how do we keep two nets from destroying each other” into “how do we predict this step’s noise more accurately.” After that rewrite, acceleration, conditioning, and latent space sit on one road.

DDPM delivered a trajectory you can inspect, not only attractive endpoints. A thousand steps are both a time tax and the operational surface itself.