Why image models are different

Everything so far in Explorer has been about models that continue text. Image generators work on a completely different principle, and almost every frustrating result makes sense once you know what it is.

Sculpture, not writing

An assistant writes left to right, choosing each next word. An image model does nothing like that.

It starts with pure noise — a rectangle of random static. Then, over several passes, it repeatedly removes whatever least resembles your description. Noise, slightly less noise, a vague arrangement of shapes, a recognisable scene, a finished image.

The whole picture is being shaped at once, at every step. There is no moment where it draws the hand and then the fingers. The fingers and the sky and the coat emerge together, out of the same fog.

What that explains

Almost every well-known limitation follows directly.

Counting. Three cats is not a quantity to the model — it is a phrase that makes the image cat-ish in a three-ish way. Nothing counts. You may get three, often two or four, sometimes one cat with an unsettling number of legs.

Text in images. Letters are shapes that must be exactly right in exactly one order. A process shaping everything simultaneously towards “looks like a sign” produces sign-shaped marks. Newer models are markedly better at this, because it was worth fixing specifically — but the underlying difficulty is structural, not a bug.

Negatives. Say no text and you may well get text. Every word you write pulls the image towards what it describes, and text is one of your words. Some tools accept a separate negative prompt, which works because it is applied as a distinct signal rather than as words in your description. Where there is no such field, describe what you do want: a blank wooden sign, not a sign with no writing.

Hands. The famous one. Hands are small, highly variable, and frequently overlapping — the hardest possible case for a process that resolves an entire image together. Also markedly improved, and still where you look first.

What this means for you

Two working habits fall out.

Describe the finished image, not the process. A weathered wooden sign, blank, morning light — a scene, not an instruction to an artist.

And expect to generate several. The randomness is not a defect to be prompted away; it is how the process works. Four images and picking one is the normal loop, not a sign you asked badly.

Try it now (6 minutes)

Ask any image generator for five red apples on a plain table, no leaves.

Count the apples. Look for leaves. Neither result is a failure of your prompt — both are the mechanism, and recognising it here will save you an hour of rephrasing later.

Check your understanding

1. A diffusion model creates an image by: Drawing elements one at a time in order Starting from noise and repeatedly removing what does not match the description Assembling pieces of images from its training data

2. Image models struggle with exact counts because: Their training images were poorly labelled Nothing in the process counts — a number is a phrase shaping the whole image Counting requires more computation than is available

3. The best way to avoid text appearing in an image: Write 'no text' in the prompt Describe what you do want — 'a blank wooden sign' Request a lower resolution

Recap

Image models start from noise and shape the whole picture at once towards your description. That single fact explains the trouble with counts, spelling, negatives and hands. Describe the finished scene rather than instructing an artist, and generate several — the variation is the mechanism, not your mistake.