The History of AI

From noise to image

Three lines of work on generating images, latent space, diffusion, and denoising, come together in the image generators of 2022.

  1. Variational autoencoder (December 20, 2013)

    The variational autoencoder learns to compress images into a space of a few variables and to generate new images from it.

  2. The first diffusion model (March 12, 2015)

    Sohl-Dickstein and colleagues propose destroying an image with noise step by step and learning to travel the path back.

  3. Score-based generative models (July 12, 2019)

    Song and Ermon generate images by removing noise at several levels with Langevin dynamics.

  4. DDPM (June 19, 2020)

    DDPM shows that the two ideas are the same and produces images as good as the best GANs.

  5. DALL·E and CLIP (January 5, 2021)

    DALL·E draws from text with a Transformer that works on the pieces of a variational autoencoder, and CLIP links images and text.

  6. DALL·E 2 (April 6, 2022)

    DALL·E 2 replaces the Transformer with a diffusion model guided by CLIP's representations.

  7. Stable Diffusion (August 22, 2022)

    Stable Diffusion runs diffusion in an autoencoder's space and releases its weights.

Open in the interactive timeline