Diffusion models and generative imagery
Modelos de difusión e imagen generativa (DALL-E 2, Stable Diffusion, Midjourney)
- Date
- 2020-2022 · confidence high
- Place
- UC Berkeley (DDPM, 2020); LMU Munich, Germany (Stable Diffusion, 2022); OpenAI, San Francisco (DALL-E 2)
- Credited to
- Jonathan Ho, Ajay Jain and Pieter Abbeel (DDPM); Robin Rombach and Björn Ommer (Stable Diffusion, with Stability AI); OpenAI (DALL-E 2); Midjourney
- People
- 👤 Björn Ommer1981–👤 Pieter Abbeel1977–👤 Robin Rombach
- Field
- Knowledge & symbols
What it is
Models that learn to reverse the addition of noise to generate photorealistic images from text, released in 2022 with open weights. They brought visual creation to everyone and later to video (Sora, 2024).
What it led to
- Generative visual AI in design, advertising and film, and the dispute with artists
- Accessible deepfakes and visual misinformation
- The generation of video, music and 3D worlds
Built on
- Generative adversarial networks (GANs)2014
- Transformer architecture2017
- Transformer and CLIP
- non-equilibrium thermodynamics
Made possible
- video generation (Sora, Veo)
- AI photo editing
- molecule design by diffusion (models)