Text-to-Image

1Words turn into meaning first

A typed sentence doesn't become a picture right away. It first goes through a step that turns the meaning of the sentence into a bundle of numbers. At this stage, phrases that are worded differently but mean roughly the same thing — a yellow field, golden farmland — land on similar values.

That's different from carrying each word over one at a time. The mood of the whole sentence gets folded into the value, and the side that builds the picture uses that value as a direction to move in. That's also why a small typo still mostly gets understood.

Writing in a language other than English still works reasonably well, but since the descriptions used in training skew heavily toward English, results can wobble depending on the phrasing. If the picture that comes back isn't quite right, rephrasing the same idea a different way often helps.

2The picture starts from blur

In the most common method today, the side that builds the picture doesn't start from a blank page — it starts from a screen full of meaningless blur. Against this, it clears the blur away a little at a time, checking in with the meaning of the sentence. After several rounds of clearing (anywhere from a handful to a few dozen, depending on the model), shapes gradually settle in among the blur.

Some services show the screen slowly coming into focus, and with this method it's actually happening step by step (though sometimes it's just a visual effect). Because the sentence gets consulted again at every single clearing step, what was written keeps exerting influence the whole way through, not just at the start.

3The same sentence gives a different picture

The blur that serves as the starting point usually gets drawn fresh every time. In tools that let you lock the number that sets that blur (the seed), the same picture can sometimes be produced again. Which spots came out a bit darker changes the very first judgment call, and that difference carries all the way through. So the same sentence gives back a picture with a different layout and different colors each time.

This trait is both a nuisance and a convenience. Landing on the exact wanted picture in one try is hard, but generating several at once and picking a favorite is easy. Making small adjustments starting from a favorite is another common approach.

4What works well and what doesn't

Landscapes, objects, and illustrations with a clear style tend to come out well. On the other side, there are still things that are hard. Lettering on signs or labels has improved a lot in recent models, but longer sentences and scripts with little training material still tend to come out mangled. Parts with a fixed count and joints, like fingers or legs, can come out wrong too, depending on the model.

Counting is a weak spot as well. Ask for five flowers and four or six often show up instead. Directions describing how several things relate to each other are shaky too — say what's on the left and what's on the right, and the placement still gets swapped.

When the wanted result isn't coming through, putting the core point up front and cutting what's unnecessary tends to work better than stretching the sentence out longer. Some services set aside a separate field just for what shouldn't appear.

5Worth knowing before you use it

Whatever leanings sit in the training material tend to show up in the finished picture. Name a profession and a similar look shows up every time, for instance. It's worth a look before using a result as-is.

A picture that closely resembles a real person or closely imitates a particular artist's style can raise likeness or copyright concerns. Whether a generated picture can be used commercially differs from service to service, so it's worth checking the terms, and marking a picture as AI-made is the safer move depending on where it's used.