Denoising Diffusion Probabilistic Models
An AI-generated review of DDPM, the 2020 paper that turned diffusion models into the leading family of generative models for images.
Primary source: Denoising Diffusion Probabilistic Models (arXiv:2006.11239)
For years, GANs produced the most realistic images, but their training was unstable and they lacked a principled likelihood. In 2020, Jonathan Ho, Ajay Jain, and Pieter Abbeel presented DDPM, showing that a model trained to undo a gradual noising process could generate images competitive with GANs on FID — with stable, likelihood-based training. The paper turned diffusion models from a niche idea into the dominant family of generative models, the engine behind Stable Diffusion, DALL·E 2, and Midjourney.
The core idea
Two processes:
- Forward process (fixed): starting from a real image x₀, add small amounts of Gaussian noise over T steps, producing a ladder x₁, …, x_T that converges to pure noise. This process is designed so each step is tractable: x_t is a simple linear combination of x₀ and noise, with a closed-form marginal.
- Reverse process (learned): train a neural network to predict and remove the noise at each step. Sampling starts from pure noise and runs the reverse process backward, step by step, to synthesize an image.
The model is trained with a simple objective: predict the noise ε added at step t (or equivalently denoise x_t to x₀). This reduces to a weighted variational bound and, in the paper's simplified form, to plain mean-squared-error between predicted and actual noise — the same kind of objective used in denoising autoencoders.
How it works
- Architecture: a U-Net with time-step embeddings (sinusoidal position embeddings added to every block), ResNet blocks, and attention at low resolutions — the template inherited by every later diffusion backbone.
- Noise schedule: a fixed linear schedule of variances β₁…β_T (T = 1,000 steps in the paper) controlling how fast signal is destroyed.
- Training: sample a random step t, add the corresponding noise, and train the U-Net to predict ε. Simple, stable, and — unlike GANs — no adversarial game.
- Sampling: run the learned reverse process for T steps. DDPM's original T=1000 sampling is slow; later work (DDIM, DPM-Solvers, latent-space diffusion) made it orders of magnitude faster.
- Results: FID of 3.17 on unconditional CIFAR-10 and 7.45 on ImageNet 64×64 at the time — competitive with GAN baselines, with a proper log-likelihood (NLL) that GANs cannot provide.
Why it matters
- Stable, scalable generative training. The noise-prediction objective avoids mode collapse and adversarial instability, making large-scale training (billions of images) feasible.
- The architecture of modern image generation. Latent diffusion (2022) moved the U-Net into a compressed latent space, enabling text-to-image models trained on LAION-scale datasets and consumer GPUs.
- A shared framework. The same DDPM machinery powers image, audio, video, and protein-structure generation (including AlphaFold-related generative efforts), with conditional variants (classifier guidance, classifier-free guidance) added in 2021-22.
- Controlled editing. Because the forward process is invertible in expectation, diffusion models excel at inpainting, super-resolution, and text-guided editing — use cases where GANs struggled.
Limitations and open questions
DDPM's original sampling is slow (thousands of network evaluations per image), and its likelihoods, while principled, trail specialized likelihood models. The paper's architecture is heavy: the U-Net is large and sampling cost scales with resolution — motivating latent-space and distillation approaches. Data requirements are substantial, and like all generative models, diffusion systems raise concerns about deepfakes, memorization of training data, and bias in generated content.
Key takeaways
- DDPM = learn to reverse a fixed Gaussian noising process; train by predicting the noise added at each step.
- Stable likelihood-based training replaced adversarial games as the default for image generation.
- The U-Net + time-embedding design became the backbone of all modern diffusion systems.
- Its successors (latent diffusion, DDIM, guidance) made the approach fast and controllable enough for mainstream products.
Want the next review in your inbox? The briefing collects them — at most one email a week. Read signup data and removal options.
Cite this report
@misc{knowyourml-ddpm-denoising-diffusion-probabilistic
title = {Denoising Diffusion Probabilistic Models},
author = {{Know Your ML}},
year = {2020},
howpublished = {\url{https://knowyourml.yunusemrevurgun.com/report.php?slug=ddpm-denoising-diffusion-probabilistic}},
note = {AI-generated review of arXiv:2006.11239. Primary sources are listed on the page.}
}
Machine Learning Decision Toolkit
Turn this review into a decision: the Machine Learning Decision Toolkit packages the evaluation scorecards, workload templates and deployment checks behind model, provider and architecture choices.
References
- Denoising Diffusion Probabilistic Models (arXiv:2006.11239) — https://arxiv.org/abs/2006.11239
Part of
Beyond the paper
How to Remain Valuable When Intelligence Becomes Cheap
This report covered one more way the machine gets smarter. The companion eBook asks the question on the other side: which human advantages stay scarce — and valuable — when intelligence itself is cheap?