SESKStuff Everyone Should Know
SESK: Stuff Everyone Should Know · sesk.info/ai-diffusion-noise-and-reverse

What two processes form the basic idea of a denoising diffusion model?

A fixed process adds noise, then a learned process reverses it.

Start playing free

  1. Sorting then hashing
  2. Noising then learned denoising ✔
  3. Encoding then deleting

Why: Training corrupts data toward a simple noise distribution and learns how to reverse that corruption. Generation begins from noise and follows the learned reverse process toward a sample.

Huh, didn't know that: For common image diffusion models, the easy reference distribution is approximately Gaussian.

The dad joke
Diffusion made a mess with noise, then learned to clean it up.
Like this card?
Play the game

References

Answer: Noising then learned denoising. Training corrupts data toward a simple noise distribution and learns how to reverse that corruption. Generation begins from noise and follows the learned reverse process toward a sample.

  1. From Denoising Diffusions to Denoising Markov Models arxiv.org/abs/2211.03595
    They work by diffusing the data distribution into a Gaussian distribution and then learning to reverse this noising process to obtain synthetic datapoints.
    Checked 2026-10-10.

Think this answer is wrong? Tap "Challenge this answer" on the card in the game and tell us why.

Test yourself in the game

Artificial IntelligenceHow do AI, machine learning and deep learning fit together?Artificial IntelligenceWhat did Alan Turing call his 1950 test of machine intelligence?Artificial IntelligenceTuring predicted that by about 2000, judges would ID the machine at most what percent of the time?

That's stuff everyone should know.