Skip to content
    Flow-alerta.cab

    Diffusion Model

    AI fundamentals & modelsDenoising Diffusion Model
    Editorially reviewed on: 2 sources
    Your reading view

    The short answer · 13 words

    In short

    Generative model that generates data by gradually removing noise from a random distribution.

    Flow-alerta.cab (2026): Diffusion Model. https://flow-alerta.cab/en/glossar/diffusion-model. Editorially reviewed on: 09 October 2026.

    How much detail would you like?

    Practical

    Relevant for generative image and media applications; test result quality, rights and effort.

    Possible example: A marketing team creates draft images and then checks product details, usage rights and eligibility for publication.

    Automated creation of advertising materials, product visualization in e-commerce and creation of social media graphics.

    At a glance

    Core
    During training, the model learns to reverse a defined noise process. In the generation, an initial distribution is converted into data in several steps. Application, presentation form and sampling method determine whether, for example, images, audio or other data are generated.
    Use
    Relevant for generative image and media applications; test result quality, rights and effort.
    Example
    Possible example: A marketing team creates draft images and then checks product details, usage rights and eligibility for publication.
    Limits
    High GPU demand for inference; complex control of fine details without specialized control modules (e.g. ControlNet).

    In context

    Generative AI

    Generative AI generates new content or data patterns based on learned distributions.

    Generative Adversarial Network (GAN)

    A GAN trains generator and discriminator in an opposing educational game.

    Midjourney

    Midjourney is a service for generative visual concepts.

    Questions about this term

    Are DDPM and diffusion model identical?

    DDPM refers to a specific family of diffusion models, not all approaches.

    How long does it take to generate?

    This depends, among other things, on the model, resolution, hardware and number of sampling steps.

    All details and background

    During training, the model learns to reverse a defined noise process. In the generation, an initial distribution is converted into data in several steps. Application, presentation form and sampling method determine whether, for example, images, audio or other data are generated.

    ELI5 – Explained simply

    Like a sculptor who exposes the finished work of art from a misshapen boulder layer by layer.

    Why it matters in practice

    Relevant for generative image and media applications; test result quality, rights and effort.

    Example

    Possible example: A marketing team creates draft images and then checks product details, usage rights and eligibility for publication.

    Pitfalls and limits

    • High GPU demand for inference; complex control of fine details without specialized control modules (e.g. ControlNet).

    Origin & history

    Sohl-Dickstein et al. (2015) described generative forward and backward diffusion. Rombach et al. published the approach of latent diffusion models in 2021/2022.

    Use in mid-sized companies

    • Automated creation of advertising materials, product visualization in e-commerce and creation of social media graphics.

    Related terms

    Matching Flow-alerta service

    Brand & digital presenceBrand & digital presence; AI introduction & accompaniment – tested use of generative images

    Matching tool / use case

    View use casesUse Case: Create image designs and review them editorially and creatively.

    Sources

    • arxiv.orgSohl-Dickstein et al. (2015), Deep Unsupervised Learning using Nonequilibrium Thermodynamics
    • arxiv.orgRombach et al. (2021/2022), High-Resolution Image Synthesis with Latent Diffusion Models

    Quellen geprüft am 09.10.2026.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases