Home/Generative AI/Diffusion Model
Generative AI

Diffusion Model

How most AI image tools work — starting from random noise and removing it step by step, guided by a prompt, until a picture appears.

Reading level: Curious
Pick your depth ↓

When not to use it

  • When you need the same output twice. Diffusion is stochastic by design. For deterministic assets, generate once and store the file.
  • For text inside images, precise counts, or exact layouts. These are known weak spots; the model is painting what text looks like, not typesetting it.
  • Where provenance matters. If you can't say where the training data or the output came from, that's a legal and editorial question before it's a technical one.

Reach for something else instead

  • Autoregressive image models and GANs each trade differently on speed, diversity, and control; GANs are still faster at inference for narrow domains.
  • Templates and design tools when you need exact, repeatable, brand-correct output — which is most commercial work.
  • Stock or commissioned imagery when licensing clarity is worth more than novelty.

Sources & further reading

  • Ho, Jain & Abbeel (2020), Denoising Diffusion Probabilistic Models — the paper that made diffusion work.
  • Rombach et al. (2022), High-Resolution Image Synthesis with Latent Diffusion Models — Stable Diffusion, and the move to latent space that made it runnable on consumer hardware.
  • Song, Meng & Ermon (2020), Denoising Diffusion Implicit Models — fewer steps, and the speed/quality trade-off you actually tune.

Primary sources, listed so you can check the claims on this page rather than take them on trust.

Where people go wrong

  • Prompting harder to fix a structural failure. If the model can't do hands or text, more adjectives won't help — inpainting or a different tool will.
  • Ignoring the step-count/quality trade-off, then complaining about latency. Fewer steps is often nearly as good and much faster.
  • Treating the seed as irrelevant. It's the one lever that makes results reproducible enough to iterate on.

At a glance

FieldGenerative AI
Core ideadenoise from static to image
Trade-offquality vs. number of steps
Key controlssteps, guidance, seed
DifficultyIntermediate
Flashcards for this concept
Question
Answer
1 / 4

Often compared with

Diffusion vs. GAN — iterative denoising vs. a generator trained against a discriminator.