MEPX
Chapter 3 of 10All chapters
  1. 1What generative means
  2. 2Text generation
  3. 3Images and other media
  4. 4Working with these tools
  5. 5What to expect it to get wrong
  6. 6Customising a model
  7. 7Tools and agents
  8. 8Copyright and licensing
  9. 9Provenance and trust
  10. 10Using it well at work

Chapter 3 of 10

Images and other media

Diffusion and its relatives.

Diffusion

Training teaches the model to reverse a process that gradually adds noise. Generation starts from pure noise and removes it step by step, guided by the prompt.

  • This is why image generation is iterative and why intermediate steps look like fog.
  • Earlier approaches used adversarial training between two competing networks.

Audio and video

The same principles extended along time. Video is harder because consistency between frames must be maintained, which is where most current artefacts appear.