0.1Generative image AI: how it works
The essentials in 30 seconds
- A diffusion model starts from random noise: trained to remove noise step by step, it makes a sharp image emerge from an initial 'snow'.
- Your prompt GUIDES the denoising: converted into embeddings, your text steers each step — so the wording strongly changes the result.
- Generation is probabilistic: same prompt ≠ same image (except with a fixed seed) — and the AI generates something new, it does not paste together pieces of existing images.
Running mini-case — Sonia tests randomness in Algiers
Sonia, freelance graphic designer in Algiers, discovers visual AI to illustrate a client's articles. She types 'a lighthouse on a cliff at sunset', generates twice — and gets two different images. Instead of crying bug, she applies the guide's method: she reruns several times to explore, then adds ONE detail at a time ('seen from afar, stormy sky') to see what each word changes. She writes the lesson on a cheat sheet: the model denoises random noise guided by her text — no magic prompt, but wording to refine.
Standalone case for this section: it returns in the course, the diagram, the script and the 48h challenge of section 0.1.
1. The diffusion model: from noise to image
A generative image AI turns a sentence (the prompt) into an image through the field's dominant principle: the diffusion model. During training, the model is shown millions of images to which noise (random 'snow') is progressively added until they become unrecognizable, and it learns the reverse path: removing the noise step by step to recover a sharp image. Once trained, it can therefore start from pure random noise and 'denoise' it until an image appears. Sonia remembers the picture: you do not paint, you reveal — like a photo developing.
2. The text guides the denoising
It remains to decide which image to make appear: that is the prompt's role. Your text is converted into a mathematical representation (embeddings) that guides the denoising at each step, so the result matches your description. The model does not 'understand' in the human sense: it steers generation toward what statistically resembles your text. That is why wording matters so much: each precise word tightens the guidance, each vague one loosens it. Sonia tests it: 'a lighthouse' gives a base, 'a lighthouse on a cliff at sunset, seen from afar' gives a direction.
3. Probabilistic, not copy-paste
Two practical consequences to remember. First, the result is probabilistic: two generations with the same prompt give different images, because each starts from different initial noise — unless you freeze chance with a seed (section 2.2). Second, the AI does not copy existing images: it produces a new image from learned regularities — which does not remove the legal and ethical questions (section 0.4). Understanding how it works avoids two beginner mistakes: believing a 'magic' prompt exists, and being surprised by variability. Sonia posts her rule: to explore, rerun; to iterate seriously, freeze the seed and change one thing at a time.
Did you know? Developing like film photography
No magic: noise, text, and an image that reveals itself.
🔎 Concrete example — 🔎 Concrete example — Variability in action
Sonia types 'a lighthouse on a cliff at sunset' and reruns twice identically.
/imagine prompt: a lighthouse on a cliff at sunset
→ image 1: centered lighthouse, calm sky.
/imagine prompt: a lighthouse on a cliff at sunset (identical rerun)
→ image 2: different framing and clouds.
/imagine prompt: a lighthouse on a cliff at sunset, seen from afar, stormy sky
→ the added detail strongly redirects the result.Reading: the first two images differ although the prompt is identical — proof that the starting noise changes each time.
Lesson: the third prompt shows the other side: wording guides — one concrete detail moves the result.
Section vocabulary
- Prompt
- Text description sent to the model; it guides the image generation.
Simple example: Sonia writes 'a lighthouse on a cliff at sunset' to launch her generation.
What it is not: a magic formula that would always give the same image.
Reflex: describe what must be seen, with concrete words.
- Diffusion model
- A model that learns to remove noise step by step; it starts from random noise to make an image emerge.
Simple example: Sonia pictures denoising as a developing photo.
What it is not: a bank of images from which pieces would be picked.
Reflex: think 'guided revelation', not 'assembly'.
- Denoising
- Iterative noise-removal process that progressively makes the image appear.
Simple example: At each step, Sonia's image becomes a little sharper.
What it is not: a simple upscaling of existing pixels.
Reflex: let the process run to the end before judging.
- Embedding
- Mathematical representation of the text that lets the model steer generation.
Simple example: Sonia's prompt becomes quantified guidance for the denoising.
What it is not: a human understanding of the meaning of words.
Reflex: craft the wording carefully: each word weighs on the guidance.
- Probabilistic generation
- Random nature of the result: the same prompt gives different images on each run.
Simple example: Sonia reruns twice and compares two different lighthouses.
What it is not: a bug or malfunction of the tool.
Reflex: to explore: rerun; to iterate: freeze the seed.
In practice — In practice — Observe the probabilistic nature of generation
- Pick an accessible tool (see sections 0.2 and 1.4) and enter a simple prompt, for example 'a lighthouse on a cliff at sunset', like Sonia.
- Generate once, then rerun EXACTLY the same prompt.
- Compare: the two images differ — this is normal, generation starts from different random noise.
- Add a detail to the prompt ('seen from afar, stormy sky') and observe how strongly wording steers the result.
Ready-to-say script — explaining generation to a client (30s)
'The AI does not copy photos: it starts from random noise and denoises it step by step, guided by my text. So the same prompt gives different images each time — that is normal, it is even useful for exploring. Like Sonia, I rerun several times to find the track, then I refine word by word.'
Say while showing the noise → denoising → image diagram; stress 'no copy-paste'.
Copy-paste template — my first variability test
/imagine prompt: a lighthouse on a cliff at sunset
→ generate, note what you like and what is missing.
/imagine prompt: a lighthouse on a cliff at sunset
→ rerun IDENTICALLY, compare: what changed on its own?
/imagine prompt: a lighthouse on a cliff at sunset, seen from afar, stormy sky
→ add ONE detail, compare: what did that word move?With Midjourney
Sonia runs the model in the Midjourney web app: a grid of four proposals per generation, perfect for comparing reruns of the same prompt.
With a free tool
With no subscription: same test via a free offer giving access to DALL·E 3 (quotas) — variability shows the same way, two reruns are enough.
Exploration notebook
Sonia pastes her three generations side by side in a document, with the exact prompt under each: her first embryo of a prompt library.
Pair in Algiers
Sonia shows two reruns to a colleague without saying the prompt is identical: he looks for the wording difference — then she reveals the chance of the starting noise.
✓ Best practices
- Rerun several times: never judge an idea on a single image.
- Word with concrete visual words: wording guides the denoising.
- Freeze the seed (section 2.2) as soon as you want to iterate seriously on a track.
✗ Common pitfalls
- Believing a 'magic' prompt exists that always gives the same result.
- Being surprised by the variability of results: it is intended, not a bug.
- Thinking the AI pastes together pieces of existing images.
Key takeaways
- A diffusion model starts from random noise and removes it step by step to make the image appear.
- The prompt, converted into embeddings, GUIDES this denoising: wording strongly changes the result.
- Generation is probabilistic: same prompt ≠ same image (except with a fixed seed).
- The AI does not paste together pieces of existing images: it generates from learned regularities.
48h challenge — apply tomorrow
Run the test model like Sonia (simple prompt + identical rerun + adding one detail), paste the three images side by side and note in one sentence what each run taught you.
- I can explain the diffusion model in one sentence (noise → guided denoising → image).
- I reran a prompt identically and observed the variability.
- I measured the effect of a single word added to the prompt.
Frequently asked questions
Does the AI copy existing images to make mine?
No, not in the copy-paste sense: the model has learned statistical regularities and GENERATES a new image by denoising a random starting point. Nuances: memorization cases exist for highly repeated images (ultra-famous artwork, logo) — a reason to check what you publish — and the 'new' nature of the result does not settle the legal question of training data (ongoing lawsuits, rights varying by country: see 0.4).
Why does the same prompt never give the same image twice?
Because each generation starts from different random noise. This is an intended property: it lets you explore several interpretations of the same idea. To iterate seriously, freeze the seed (same prompt, same version, same parameters = reproducible result, section 2.2) and change only one thing at a time.
How does a diffusion model create an image?