Denoising Diffusion Probabilistic Models

The paper made diffusion models a practical high-quality generative modeling recipe by training a neural network to reverse a fixed gradual noising process.

Metadata

Field	Value
Paper	Denoising Diffusion Probabilistic Models
Authors	Jonathan Ho, Ajay Jain, Pieter Abbeel
Year	2020
Venue	NeurIPS 2020
arXiv	2006.11239
Project page	hojonathanho.github.io/diffusion
Code	hojonathanho/diffusion
Status	full note started

One-Line Takeaway

DDPM turns generation into iterative denoising: corrupt data with a known Gaussian Markov chain, train a neural network to reverse each step, and sample by starting from noise and repeatedly denoising.

Question

Before diffusion became dominant, image generation was often read through:

• GANs for sharp samples;
• VAEs for likelihood-based latent-variable modeling;
• autoregressive models for tractable likelihood and sequential decoding;
• normalizing flows for exact invertible density models.

DDPM asks:

Can a latent-variable Markov chain trained with a variational objective produce high-quality samples by reversing a simple noising process?

The generative contract:

$$x_T\sim \mathcal{N}(0,I),x_{T-1},\dots ,x_0\sim p_{\theta }(x_{t-1}\mid x_t).$$

The sample is $x_0$.

Main Claim

The narrowed architecture/objective claim:

$$\text{fixed Gaussian noising}+\text{learned denoising network}+\text{weighted variational objective}\Rightarrow \text{high-quality image synthesis}.$$

The paper is important because it made a clean recipe:

1. define a forward diffusion process $q$ that gradually destroys data;
2. train a reverse process $p_{\theta }$ to undo it;
3. use a simplified denoising objective;
4. sample through many reverse steps.

Architecture Contract

Component	Role
data $x_0$	clean image or sample
forward process $q$	fixed Gaussian noising chain
reverse process $p_{\theta }$	learned denoising chain
denoising network	predicts reverse transition statistics or noise
timestep embedding	tells the network which noise level it is denoising
noise schedule	controls corruption strength over time
sampling loop	iteratively maps Gaussian noise to data

The model is an architecture family plus an objective, not a single network block. In images, the denoising network is often a U-Net-like architecture, but the DDPM idea can be used with other denoisers and domains.

Forward Diffusion

The forward process gradually adds Gaussian noise:

$$q(x_t\mid x_{t-1})=\mathcal{N}\left(x_t;\sqrt{1-{\beta }_t}{x}_{t-1},{\beta }_{t}I\right),$$

where:

• $t\in \{1,\dots ,T\}$;
• ${\beta }_t$ is the variance schedule;
• small ${\beta }_t$ means a small noising step.

Define:

$${\alpha }_t=1-{\beta }_t,{\bar{\alpha }}_t=\prod _{s=1}^t{\alpha }_s.$$

Because the forward chain is Gaussian, there is a closed-form marginal:

$$q(x_t\mid x_0)=\mathcal{N}\left(x_t;\sqrt{{\bar{\alpha }}_t}{x}_0,(1-{\bar{\alpha }}_t)I\right).$$

This allows one-step sampling of $x_t$ from $x_0$:

$$x_t=\sqrt{{\bar{\alpha }}_t}{x}_0+\sqrt{1-{\bar{\alpha }}_t}ϵ,ϵ\sim \mathcal{N}(0,I).$$

This equation is the practical training workhorse.

Reverse Process

The learned reverse process is:

$$p_{\theta }(x_{t-1}\mid x_t)=\mathcal{N}\left(x_{t-1};{\mu }_{\theta }(x_t,t),{\Sigma }_{\theta }(x_t,t)\right).$$

Sampling starts from:

$$p(x_T)=\mathcal{N}(0,I).$$

Then:

$$x_T\to x_{T-1}\to \cdots \to x_0.$$

The model has to learn local denoising transitions, not a direct one-shot map:

$$z\to x.$$

That is the main generative architecture shift.

Noise Prediction Parameterization

The paper shows a useful parameterization where the network predicts the noise $ϵ$ used to corrupt $x_0$:

$${ϵ}_{\theta }(x_t,t)\approx ϵ.$$

Given:

$$x_t=\sqrt{{\bar{\alpha }}_t}{x}_0+\sqrt{1-{\bar{\alpha }}_t}ϵ,$$

we can estimate:

$${\hat{x}}_0=\frac{x_t-\sqrt{1-{\bar{\alpha }}_t}{ϵ}_{\theta }(x_t,t)}{\sqrt{{\bar{\alpha }}_t}}.$$

The learned mean can be written in terms of predicted noise:

$${\mu }_{\theta }(x_t,t)=\frac{1}{\sqrt{{\alpha }_t}}\left(x_t-\frac{{\beta }_t}{\sqrt{1-{\bar{\alpha }}_t}}{ϵ}_{\theta }(x_t,t)\right).$$

This form is why noise prediction became a standard diffusion interface.

Simplified Training Objective

The simplified DDPM loss is:

$${\mathcal{L}}_{\text{simple}}={\mathbb{E}}_{x_0,t,ϵ}\left[{\Vert ϵ-{ϵ}_{\theta }\left(\sqrt{{\bar{\alpha }}_t}{x}_0+\sqrt{1-{\bar{\alpha }}_t}ϵ,t\right)\Vert }_2^2\right].$$

Reading:

1. sample a real data point $x_0$;
2. sample a timestep $t$;
3. sample Gaussian noise $ϵ$;
4. construct noisy $x_t$;
5. train the network to recover $ϵ$.

The model is not trained by running the full chain during every training example. The closed-form noising equation makes training efficient.

Variational Bound Reading

DDPM is a latent-variable model:

$$p_{\theta }(x_{0:T})=p(x_T)\prod _{t=1}^T{p}_{\theta }(x_{t-1}\mid x_t).$$

The forward process defines an approximate posterior:

$$q(x_{1:T}\mid x_0)=\prod _{t=1}^{T}q(x_t\mid x_{t-1}).$$

The evidence lower bound can be written as:

$$\log p_{\theta }(x_0)\ge {\mathbb{E}}_q\left[\log \frac{p_{\theta }(x_{0:T})}{q(x_{1:T}\mid x_0)}\right].$$

The DDPM contribution is partly that this variational view connects to denoising score matching and yields a simple noise-prediction loss that works well in practice.

Score Matching Connection

The score of a noisy distribution is:

$${\nabla }_{x_t}\log q(x_t).$$

Under Gaussian corruption, predicting noise is related to predicting the score:

$$s_{\theta }(x_t,t)\approx -\frac{{ϵ}_{\theta }(x_t,t)}{\sqrt{1-{\bar{\alpha }}_t}}.$$

This connects DDPM to score-based generative modeling:

$$\text{denoise}\leftrightarrow \text{estimate score}\leftrightarrow \text{reverse stochastic dynamics}.$$

The paper is therefore a bridge between variational latent-variable modeling and score-based sampling.

Sampling Algorithm

A simplified DDPM sampler:

1. sample $x_T\sim \mathcal{N}(0,I)$;
2. for $t=T,\dots ,1$:
3. predict ${ϵ}_{\theta }(x_t,t)$;
4. compute ${\mu }_{\theta }(x_t,t)$;
5. sample $x_{t-1}$ from the reverse Gaussian;
6. return $x_0$.

Mathematically:

$$x_{t-1}={\mu }_{\theta }(x_t,t)+{\sigma }_{t}z,z\sim \mathcal{N}(0,I).$$

For $t=1$, implementations may omit extra noise.

This iterative sampler is the source of both:

• high sample quality;
• slow sampling relative to one-shot generators.

Progressive Denoising

DDPM generation can be viewed as coarse-to-fine reconstruction:

$$\text{noise}\to \text{coarse structure}\to \text{details}\to \text{sample}.$$

The paper also frames the process as progressive lossy decompression. Early reverse steps recover broad semantic structure; later steps recover fine detail.

For later architectures, this view becomes important:

• image diffusion uses U-Net multi-scale denoisers;
• latent diffusion denoises compressed latent variables;
• video diffusion extends denoising over time;
• molecular diffusion denoises coordinates, graphs, or atom types.

Relation to GANs

Axis	GAN	DDPM
generator	one-shot map from latent to sample	iterative denoising chain
training signal	adversarial discriminator	denoising/variational objective
likelihood	usually not tractable	variational bound available
sampling speed	fast	often slow
sample quality	sharp but unstable training risk	high quality with stable denoising training
mode coverage	can suffer mode collapse	generally better coverage behavior

DDPM did not make GANs irrelevant immediately. It changed the tradeoff: slower sampling, but stable training and strong sample quality.

Relation to VAEs

DDPM is a latent-variable model, but its latent variables are a long noising trajectory:

$$x_1,\dots ,x_T.$$

VAE-style model:

$$z\sim p(z),x\sim p_{\theta }(x\mid z).$$

DDPM-style model:

$$x_T\sim \mathcal{N}(0,I),x_{t-1}\sim p_{\theta }(x_{t-1}\mid x_t).$$

The decoder is not one big conditional distribution. It is many small denoising steps.

Relation to Autoregressive Models

Autoregressive models factorize data by order:

$$p(x)=\prod _{i}p(x_i\mid x_{<i}).$$

DDPM factorizes generation by noise level:

$$p_{\theta }(x_0)=\int p(x_T)\prod _{t=1}^T{p}_{\theta }(x_{t-1}\mid x_t)d{x}_{1:T}.$$

The decoding order is not pixel-by-pixel or token-by-token. It is denoising-time order.

Relation to Flow Matching

DDPM learns stochastic reverse transitions:

$$x_t\to x_{t-1}.$$

Flow matching and rectified flow often learn a velocity field:

$$\frac{d{x}_t}{dt}=v_{\theta }(x_t,t).$$

Both are ways to transform a simple base distribution into a data distribution:

$$\mathcal{N}(0,I)\to p_{\text{data}}.$$

DDPM is the historical anchor for reading later diffusion/flow papers. Always ask:

• Is the process stochastic or deterministic?
• Is the target noise, score, clean data, or velocity?
• Is the sampler SDE-like, ODE-like, or discrete Markov?
• How many steps are used at evaluation?

Relation to AlphaFold3 and Molecular Models

AlphaFold3 uses a diffusion-based architecture for biomolecular coordinate generation. The object differs, but the generative idea is recognizable:

$$\text{noisy coordinates}+\text{conditioning context}\to \text{less noisy coordinates}.$$

For molecular/protein models, DDPM provides the general reading vocabulary:

DDPM Concept	Molecular/Structure Analogue
$x_0$ image	molecule, conformation, protein complex, coordinates
Gaussian noising	coordinate noise, graph corruption, categorical noise
denoising network	equivariant GNN, Transformer, structure module
timestep	noise level
condition	pocket, sequence, scaffold, text, target
sample validity	chemical/geometric validity

This is why DDPM belongs in architecture notes, not only generative-model notes.

Evidence Reading

The paper reports high-quality image synthesis, including strong CIFAR-10 FID and LSUN sample quality.

Evidence should be read by claim:

Evidence	Supports	Does Not Prove Alone
CIFAR-10 FID	strong sample quality for small images	universal superiority across domains
LSUN samples	scalability to higher-resolution image domains	fast sampling
variational bound	likelihood-oriented interpretation	best likelihood among all models
rate-distortion curves	progressive compression interpretation	semantic controllability
sample visual quality	practical generative usefulness	calibrated probability or downstream utility

The central evidence is that diffusion models can be competitive high-quality generators, not just theoretical latent-variable models.

Implementation Notes

Implementation details that change behavior:

Choice	Why It Matters
noise schedule ${\beta }_t$	controls signal-to-noise progression
timestep embedding	lets denoiser condition on noise level
denoiser architecture	controls inductive bias and capacity
parameterization	noise, clean data, score, or velocity target
variance handling	affects likelihood and sampling
number of sampling steps	quality/latency tradeoff
data scaling	Gaussian assumptions depend on normalized data

For images, the denoiser is often U-Net-like:

$${ϵ}_{\theta }(x_t,t)={\mathrm{UNet}}_{\theta }(x_t,\mathrm{embed}(t)).$$

For graphs or 3D structures, the denoiser should match the object symmetry and representation.

Failure Modes

Failure Mode	Why It Matters
ignoring sampler step count	quality and latency depend heavily on it
comparing FID without sampling budget	unfair if one model uses many more evaluations
treating diffusion as one architecture	denoiser architecture still matters
confusing noise prediction with score prediction conventions	scale and sign depend on parameterization
assuming likelihood and sample quality always align	they can trade off
porting image DDPM directly to molecules	chemical validity and symmetry need domain-specific design
ignoring conditioning leakage	conditional generation can use unavailable information

Common Misreadings

”DDPM is just adding noise and removing it.”

That is the intuition, but the paper’s value is the probabilistic formulation, the closed-form noising marginal, the reverse Markov chain, and the simplified denoising objective.

”The denoising network is the whole architecture.”

No. The architecture includes the noising process, timestep parameterization, reverse transition parameterization, denoising network, and sampler.

”Diffusion models are always slow.”

Original DDPM sampling is slow because it uses many denoising steps. Later samplers and parameterizations reduce the step count, but the speed-quality tradeoff remains a core reading axis.

”DDPM only matters for images.”

No. The image results made it visible, but the diffusion recipe became a general template for audio, video, 3D, molecules, proteins, and control.

Later-Paper Checklist

When reading later diffusion or flow papers, check:

• What is the corrupted object: pixels, latents, tokens, graphs, coordinates, fields?
• Is the process discrete-time or continuous-time?
• Does the model predict noise, score, clean data, or velocity?
• What is the noising schedule?
• What denoiser architecture is used?
• Is the sampler stochastic or deterministic?
• How many sampling steps are reported?
• Is guidance used?
• Is evaluation budget matched?
• Are invalid samples filtered before metrics?
• Does the model preserve domain symmetries such as permutation, rotation, or translation?

Why It Matters

DDPM is a foundation paper for modern generative modeling because it turned a conceptually simple denoising process into a competitive generative architecture.

For this wiki, it connects:

$$\text{latent variable models}\to \text{score matching}\to \text{iterative denoising}\to \text{modern diffusion and flow models}.$$

It also gives the vocabulary needed to read diffusion-based scientific architectures such as AlphaFold3, molecular conformation generation, and protein design models.

Limitations

The original DDPM recipe has important limitations:

• sampling requires many steps;
• image denoisers do not automatically transfer to structured scientific objects;
• sample quality depends strongly on schedule and architecture;
• likelihood and perceptual quality are not identical;
• evaluation metrics such as FID are imperfect;
• conditional generation and guidance require separate analysis.

The defensible claim:

$$\text{DDPM}\Rightarrow \text{diffusion as a practical high-quality generative model family}.$$

The overclaim to avoid:

$$\text{DDPM}\Rightarrow \text{all generative modeling problems are solved}.$$

Connections

• Diffusion model
• Score-based model
• Score matching
• ELBO
• Sampling
• Flow matching
• Rectified flow
• Probability flow ODE
• GAN
• VAE
• Autoregressive model
• Conditional generation
• 3D structure
• U-Net
• AlphaFold3
• Generative Model Papers
• Architecture papers