AI Image Generator From Text: How It Actually Works

You write a sentence, you get a picture. That part works. What surprises most people is the second image — same prompt, different face — and the discovery that on a companion platform, the thing you’re actually paying for isn’t image quality at all. It’s whether the character stays the same person.

What the machine is actually doing

A text-to-image generator doesn’t search for pictures and it doesn’t assemble them from pieces. It starts from visual noise and removes it, step by step, steering each step toward something that matches your description. The output is new every time — which is why nothing is ever retrieved twice, and why the same prompt run twice gives you two different results.

That single fact explains most beginner frustration. People expect a picture to have been found. It was generated, and generation is probabilistic by design.

Why the second image doesn’t look like the first

This is the gap between a general image tool and a companion platform, and it’s worth understanding before you judge either.

A standalone generator has no reason to remember anything. Each request is independent, so “a woman with dark hair” produces a different woman each time. On a companion platform, the character is supposed to persist — same face, same build, across weeks of conversation. Doing that requires holding a stable visual identity and steering every generation back toward it.

Platforms solve it with varying success, and it’s the single clearest quality difference in the category. A platform whose images are individually beautiful but never the same person is worse, for this purpose, than one producing plainer images of someone recognisable.

Prompting, minus the folklore

There’s a substantial folklore around prompts — long incantations of quality words, borrowed keyword lists, tags copied between tools that don’t share an architecture. Most of it does less than people believe.

What consistently matters:

• Subject before styling. Who or what, then what they’re doing, then where. Styling words added before the subject is established tend to pull the whole image toward an aesthetic and away from your description.
• Concrete over evocative. “Sitting on a windowsill at night, city lights behind” gives the model something to place. “Moody, ethereal, dreamlike” gives it a mood and no scene.
• One change at a time. Rewriting five things at once means you learn nothing about which one worked.
• Negatives, where supported. Naming what you don’t want is often faster than describing what you do.

And what doesn’t transfer: prompt syntax is specific to the tool. Weight notation from one interface is frequently ignored by another, and copying a long prompt from a forum into a companion platform usually produces something worse than a plain sentence would.

The part nobody mentions in the marketing

Images are the most expensive thing these platforms do, and they’re almost never included in the subscription in any meaningful quantity. The usual arrangement is a monthly fee for unlimited text, plus a credit balance for images — and images are what people run out of.

It matters more than it sounds, because the failed generations cost the same as the good ones. A platform that takes four attempts to give you a usable picture is four times as expensive as its pricing page implies, at identical advertised rates. We cover the mechanism in what an AI companion really costs.

Where the limits come from

Every generator refuses some requests, and the refusals come from four different places that look identical from the outside: the model’s own training, a moderation layer running alongside it, the platform’s terms, and the law of wherever the platform operates.

Knowing which one you’ve hit is genuinely useful. A moderation layer produces inconsistent refusals — the same request works one day and fails the next. Model-level limits are stable. Legal limits are absolute and identical everywhere. If your results vary day to day, you’re meeting a moderation layer, not a boundary. We go through this in the NSFW image generation guide.

How we evaluate it

Image generation is one of our six weighted axes, and we score three things: whether the image matches the character as described, whether it matches the same character across repeated generations, and what a usable result actually costs in credits — counting the failures, because you paid for those too.

We don’t score aesthetics. Taste isn’t measurable and every platform’s showcase gallery is its best work, not its typical work. Consistency and cost are measurable, and they’re what determine whether you’re still using the feature in a month.

Read the full protocol · Making AI art safely · What “free” actually gets you