How does a picture come out of noise?

1 Noise to picture

Start from noise; clean every pixel a little each step

Song and colleagues (2021) describe a diffusion model as noise added gradually and then removed in reverse: “Creating noise from data is easy; creating data from noise is generative modeling.”

Source, checked 24 October 2026

2 Step by step

Fewer steps cost less and draw rougher

Ho, Jain and Abbeel’s 2020 paper (DDPM) made high-quality pictures with a diffusion model. It set T = 1,000 steps for all its experiments, so sampling one picture runs the network 1,000 times.

Source, checked 24 October 2026

3 A learned denoiser

A network learns which way to move each dot

4 Copy or create

The ideal denoiser copies; a network makes new points

Carlini and colleagues (2023) showed that diffusion models memorise individual training images and emit them when generating, extracting over a thousand training examples from state-of-the-art models.

Source, checked 24 October 2026

5 Steer with a label

Guidance pushes toward the part asked for

Classifier-free guidance (Ho and Salimans, 2022) trains one model with and without the label, dropping it at random, and mixes the two predictions as (1 + w) × labelled − w × unlabelled, trading sample diversity for quality.

Source, checked 24 October 2026

6 Why it costs

Steps × positions; latent and few-step models cut each

Latent diffusion (Rombach et al., 2022) runs the denoiser on a pretrained autoencoder’s smaller latent picture, shrunk by a factor f each way; the paper found f of 4 to 16 a good balance, with 4 and 8 best.

Source, checked 24 October 2026