What the model is doing
Your prompt is encoded into a numeric representation of meaning. The model then starts from a field of random noise and repeatedly removes a little of it, at each step nudging the picture toward something that would plausibly be described by that representation. After a few dozen steps the noise has resolved into an image. Some newer systems predict image tokens in sequence instead of denoising, but the practical consequence is identical: you are drawing a sample from a distribution, not looking anything up.
That is why identical words can give completely different pictures, and why "the model got it wrong" is often better read as "the model gave you one of many valid answers". Rerunning is not a workaround, it is how the tool is meant to be used.
Where your control actually sits
- The prompt. Order matters. Earlier tokens carry more weight, so lead with what you are least willing to lose.
- The seed. The one lever that makes a result repeatable. Note the seed of anything you might want to revisit.
- Aspect ratio. Not a crop. A 16:9 request and a 2:3 request from the same prompt produce different compositions, because the model was trained on differently framed pictures.
- Reference images. The moment you add one you have left this task and moved into image to image.
- Negative prompts. Useful for recurring artifacts, useless as a substitute for control the model does not have.
How to compare text to image models
Six axes, and almost no model leads on all of them:
| Axis | What to test with |
|---|---|
| Prompt adherence | A deliberately over-specified prompt, then count what survived |
| Text rendering | One short quoted string on a sign or label |
| Anatomy | Hands doing something, not hands at rest |
| Style range | The same subject as photo, ink drawing, and 3D render |
| Aspect and resolution | Your actual delivery ratio, not a square |
| Cost and latency | How many attempts you can afford per idea |
Cost belongs on that list because output variance is high by design. A model you can run eight times often beats a better model you can run twice.
A prompt structure that holds up
[subject] [doing what] in [environment], [lighting], [lens and framing], [medium or style]
Keep it to one subject. With two subjects, attribute binding starts failing: ask for a woman in a red coat and a man in a blue coat and you will regularly get the colors swapped, because the model has no reliable mechanism tying each attribute to each person. If you need two specific subjects, make them separately and combine with an editing pass.
Three habits that waste tokens: stacking quality adjectives (ultra detailed, 8k, masterpiece) which the model mostly ignores now; giving contradictory framing (a wide establishing shot that is also a tight portrait); and phrasing exclusions positively (writing "no hands" tends to summon hands, because the words are present either way).
The failure modes worth recognizing
Garbled lettering, miscounting anything above three, attribute bleed between subjects, unwanted symmetry, and style collapse toward the model's house look are all normal and none of them are fixed by writing more. Each has a specific answer: composite real type, split the subjects, change the seed, or pick a model whose default aesthetic is closer to what you want.