00 · IN THREE MINUTES

The answer in three steps

  1. 1Training repeatedly adds noise to examples and teaches a network to estimate what noise was added.
  2. 2Generation starts from random noise and removes predicted noise through many steps.
  3. 3Text guidance steers the denoising path; it does not retrieve and paste a single training image.

01 · TRAINING BEGINS WITH DESTRUCTION

Training begins with destruction

Known images are corrupted by controlled amounts of random noise. A neural network is trained to predict the noise, or an equivalent quantity, using the noisy image and a timestep as inputs.

02 · GENERATION REVERSES THE SCHEDULE

Generation reverses the schedule

At sampling time there is no original image to recover. The process begins with fresh noise and repeatedly applies learned estimates, gradually creating edges, regions, textures and objects that fit the model’s distribution.

FIG. 02How the system changes state
A conceptual mechanism map. Geometry, scale and timing are explanatory unless labelled otherwise.

03 · TEXT CHANGES THE DIRECTION

Text changes the direction

A text encoder maps the prompt to a representation used during denoising. Guidance strengthens changes that make the evolving image more compatible with the prompt, while random initialization allows many outputs from the same words.

04 · LATENT DIFFUSION COMPRESSES THE CANVAS

Latent diffusion compresses the canvas

Many modern systems denoise a lower-dimensional latent representation rather than every pixel directly. A decoder turns the final latent into pixels, reducing computation while preserving much visual structure.

05 · GENERATION IS NOT EVIDENCE

Generation is not evidence

The output reflects statistical regularities, prompt constraints and sampling choices. It can reproduce biases, make physically impossible details or occasionally memorize examples; photorealism does not turn it into a photograph.

06 · SOURCES AND EVIDENCE

Sources and evidence

Claims are linked to foundational papers, standards or the primary study behind the update.

  1. 01
    Denoising Diffusion Probabilistic Models

    Supports a defined mechanism, measurement or evidence boundary in this article.

    PRIMARY STUDY ↗
  2. 02
    High-Resolution Image Synthesis with Latent Diffusion Models

    Supports a defined mechanism, measurement or evidence boundary in this article.

    PRIMARY STUDY ↗
CHANGE LOG29 Aug 2026 · First five-language edition; mechanisms, limits, diagrams and sources checked.