Generative image models (GANs to diffusion)

Neural networks that synthesise realistic images from noise or text prompts, transforming illustration, design and raising deepfake concerns.

AI & roboticsSoftware
  1. 2014

    First version

    Ian Goodfellow et al. (Université de Montréal) — GANs; Jonathan Ho et al. — diffusion

    Image: Université de Montréal (logo designed by Jean-Claude Lahaye) · Public domain (Commons) · Wikimedia Commons
  2. 2018

    StyleGAN (NVIDIA)

    Photorealistic, controllable synthetic faces.

    Image: Nvidia · Apache License 2.0 · Wikimedia Commons
  3. 2020
    Denoising diffusion (DDPM)

    Denoising diffusion (DDPM)

    Iterative denoising beats GANs on quality and stability.

    Image: Google, via its Google I/O 2018 video · Copyright Google, all rights reserved · Data Center Frontier; image from Google I/O
  4. 2022

    DALL·E 2, Midjourney, Stable Diffusion

    Text-to-image goes mainstream; latent diffusion released openly.

    Image: Original: Unknown authorUnknown author Vector: Aerra Carnicom · Public domain · Wikimedia Commons
  5. 2025
    Native multimodal image generation

    Native multimodal image generation

    GPT-4o and Gemini 2.5 Flash Image generate and edit images inside chat models.

    Image: ChatGPT (AI generation) · Public domain · Wikimedia Commons

Contested: Diffusion models were introduced by Sohl-Dickstein et al. (2015) but made practical by Ho, Jain and Abbeel (2020); Schmidhuber claims GAN precursors.

See it on the timeline