What Actually Happens When You Hit Generate
Almost every frustration artists have with these tools — the ignored instruction, the sixth finger, the four petals when you asked for five, the two wildly different images from the same prompt — comes from one thing: not knowing what the machine is doing with your sentence.
So before any prompting advice, here is the whole journey your description takes. It is four steps, and none of them require any technical background.
you type a description
|
1. a language model reads it -> turns words into meaning
|
2. the meaning becomes a target -> a 'this is what good looks like' signal
|
3. a painter starts from static -> and refines toward that target, 20-50 times
|
4. the result is decoded to pixels -> your image
The first thing that touches your prompt is a language model. This is the part of the system people mean when they say "LLM," and it is worth understanding because it is the part you actually talk to.
What a language model is, for an artist
A language model is a program that has read a colossal amount of text and learned which words tend to go with which other words, in which contexts. That's genuinely it. From that one ability it gets a surprisingly deep working knowledge of what things mean — that cerulean is a blue, that chiaroscuro implies a hard light source and deep shadow, that a Dutch angle is a tilted camera, that Bauhaus and Art Nouveau are not the same century.
It is not looking anything up. It has no dictionary. It has a learned sense of how these words sit next to each other, and that turns out to be enough.
What it does with your prompt
It converts your sentence into a compact numerical summary of its meaning — not a list of your words, but a representation of what you were getting at. Two prompts that mean nearly the same thing land in nearly the same place; two that mean different things land far apart.
This is why "a decrepit seaside hotel" and "a run-down hotel by the sea" give you similar images, while changing one adjective from decrepit to opulent changes everything.
Two flavours you'll meet in the wild
- A plain text encoder. Older and open-weight image tools use a relatively small model that just reads your words. It is literal. It does not reason, it does not fill gaps, and it does not ask what you meant. Give it a vague prompt and you get a vague image.
- A full conversational model. Tools built on chat assistants (ChatGPT's image generation, Gemini's, Copilot's) run your request through a proper reasoning model first. It will expand "a cosy reading nook" into a paragraph of specifics before any painting starts. Convenient, but it means the model is making creative decisions on your behalf — which is why the same prompt twice can give you two quite different rooms.
Knowing which of these you're talking to is the single most useful thing you can know about a tool. It decides how you should write.
That numerical summary of your meaning now becomes a target — a standing instruction that will be consulted at every single step of the painting process.
The useful mental picture is an art director standing behind the painter, holding your brief, and at every stage saying "more like this, less like that."
The director does not paint. It only judges direction. And critically, it judges the whole picture at once against the whole brief at once — it does not tick off your requirements one by one.
That single fact explains a lot of behaviour you have probably found maddening, and we'll come back to it two sections from now.
How hard the director pushes is usually a setting you control. Different tools call it different things — guidance, prompt strength, stylize, CFG, creativity — but it is always the same dial:
- Push too gently → the model wanders off and makes something pretty that isn't what you asked for
- Push too hard → the model over-corrects: scorched colours, crunchy contrast, stiff and literal compositions
- The sweet spot is usually near the default. When results look burnt or over-saturated, this dial is the first thing to turn down, not your prompt.
Here is the part that surprises people. The model does not draw your image. It does not start with a blank canvas and add a horizon, then a boat, then a sky. It starts with a rectangle of pure random static — visual noise, like an untuned television — and repeatedly *removes* the parts that don't belong.
The sculpture analogy is the accurate one. The model is not painting; it is carving. It looks at the noise and asks "if this static were a slightly blurry version of the image described in the brief, what would I need to take away?" It removes a bit. It looks again. It removes a bit more.
Do that twenty to fifty times, checking against your brief every time, and a coherent image emerges from the static. Early steps decide the big things — composition, masses, where the light comes from. Late steps decide texture, edges, fine detail.
Why this is worth knowing:
- The composition is locked in early. By roughly a quarter of the way through, the layout is decided. This is why prompt words about composition and framing carry more weight than words about small details — and why adding "tiny brass buttons" rarely gets you tiny brass buttons.
- More steps is not more quality. Past a point the image stops changing meaningfully and you're just spending money. Defaults are almost always fine; doubling the steps is not a fix for a weak prompt.
- The starting static is a real ingredient. That random rectangle is generated from a number called the seed. Same prompt + same seed + same settings = the same image, every time. Same prompt + different seed = a genuinely different image. When you get something almost right, locking the seed and changing one word is how you make a controlled adjustment instead of rolling the dice again.
For speed, most of that carving happens on a small, compressed sketch of the image rather than on the full-resolution picture — think of it as working on a thumbnail. A final decoding step expands that sketch into the actual pixels you download.
You never see this step, but it leaves fingerprints, and recognising them saves you from chasing prompt fixes for problems your prompt didn't cause:
- Small faces and hands in wide shots come out mushy. In the compressed sketch a distant face is only a few marks wide; there simply wasn't room for the detail. The fix is composition or a face-aware upscale — never a longer prompt.
- Fine repeating patterns wobble. Chain-link, houndstooth, brickwork, crowd scenes, distant text.
- This is the natural resolution ceiling. Every tool has a size it was built for. Ask for something far larger in one shot and you get duplicated limbs and repeated horizons. Generate at the native size, then upscale.
Now the payoff. Nearly every classic complaint about AI image tools is a direct, predictable consequence of the four steps above — which means each has a real fix rather than a superstition.
| What you see | Why it happens | What actually fixes it |
|---|---|---|
| "No hats" produced a hat | The brief is matched as a whole, not read as an instruction list. Hat is in the brief, so hattishness scores well. The director never learned the word "no". | Describe the positive: bare-headed, windswept hair. Or use the tool's dedicated negative-prompt field, which is a separate mechanism from your sentence. |
| You asked for five, got four | Nothing in the process counts. It matches the look of a described scene, and five petals and four petals look almost identical to a whole-image judgement. | Ask for small numbers, or arrangements it has seen often (a pair, a row of three). For exact counts, generate then edit. |
| Hands and teeth | Hands appear in a million poses, from a million angles, half-occluded. There is no single "look of a hand" to converge on — unlike a face, which is always roughly the same arrangement. | Genuinely much better in 2026 models. Otherwise: crop them out, or inpaint them afterwards. |
| It merged two subjects | The brief is one summary of one meaning. "A woman in a red coat and a man in a blue coat" is, to that summary, largely people-in-coloured-coats. Attributes bleed. | Separate them in space (on the left… on the right…), or generate separately and compose. Some tools support per-region prompting. |
| The style words did nothing | A style name only works if that style was strongly represented, and named that way, in what the model learned. Obscure or very recent artists often simply aren't in there. | Describe the style's mechanics instead of naming it: the medium, the mark-making, the palette, the light, the era. This works on every model. |
| Same prompt, wildly different images | Different starting static — and, on chat-based tools, a reasoning model reinterpreting your request afresh each time. | Lock the seed. On chat tools, be specific enough that there is nothing left to reinterpret. |
| Text came out as gibberish | The model carves shapes, and letterforms are shapes that have to be exactly right to read as language. | Use a model built for it (Ideogram, Qwen-Image, GPT Image, Recraft), keep the text short, and put it in quotation marks. |
The general principle: when a result is wrong, ask which of the four steps failed? Wrong subject or wrong meaning is step 1 — rewrite. Right subject but ignored or over-cooked is step 2 — adjust guidance. Wrong composition is step 3 — reroll or change the seed. Mushy detail is step 4 — recompose or upscale. That single question will save you more time than any prompt template.
Once you have the four steps, the other media are small variations rather than new subjects. You do not need to learn them from scratch.
- Video does the same carving, but on a stack of frames at once rather than one picture — which is why a model can keep a subject consistent as it moves. It is also why video is expensive, why clips are short, and why the thing to describe is motion and camera ("slow push in", "she turns to look over her shoulder") rather than a static scene. A described scene with no described motion gets you a slow drift.
- 3D mostly works by generating a subject from many viewpoints and reconciling them into a solid shape with a surface. That reconciling stage is why AI 3D output needs cleanup: the geometry is plausible from outside but rarely tidy underneath.
- Music and sound carve a picture of the sound over time — effectively a spectrogram — then convert it to audio. Same process, different canvas. Which is why music tools respond to the same kind of description you'd give an image: genre, instrumentation, mood, tempo, production era.
The practical upshot: everything you learn about prompting images transfers. Be concrete, describe what can be perceived, put the important thing first, and change one variable at a time.