The Button That Rewrites Your Prompt, and When To Leave It Off

Here is something most image generators do not tell you: almost none of them send your prompt to the model. They send a rewritten version. Your twelve words go to a language model first, come back as sixty, and those sixty are what the picture gets made from.

That has always happened here too. What changed this week is that it is now a switch you control, sitting just under the prompt box, off by default, marked +10 coins.

What a rewrite actually adds

Take a prompt with nothing in it: a woman in a red dress. Six words, and five of them are doing nothing. An image model has to decide the time of day, the room, the distance from the camera, the lens, whether the dress is silk or cotton, whether she is looking at you. It will decide all of that regardless. The only question is whether it decides from your words or from the statistical middle of everything it has ever seen.

A rewrite fills that space. It adds the vocabulary that image models are unusually responsive to: soft window light, shallow depth of field, the weight and texture of a fabric, where the subject is looking. None of it is magic. It is the same scaffolding described in our guide to writing prompts, written for you instead of by you.

Which is exactly why it can hurt

Now take the opposite case. You have spent twenty minutes on a prompt. You named the lens. You said the light comes from a window on the left, late afternoon, and that the fabric is heavy linen with visible weave. You specified that she is looking away from the camera.

A rewrite has a hard job with that prompt. It has to keep every constraint you set while adding its own, and the more you specified, the more there is to accidentally reword. The clause you cared about most is the one at risk, because it is often the unusual one, and unusual phrasing is what a rewrite smooths out. You do not get to see what it did. You just get a picture where the light is coming from somewhere else and no explanation why.

This is the real cost, and it is not the 10 coins. It is that you have added a variable you cannot inspect. When a render comes back wrong, you no longer know whether your prompt was wrong or the rewrite was.

The way we would use it

Treat it as a teacher rather than a setting.

When an idea is still one line in your head, turn it on. Spend the 10 coins, look at what comes back, and read the picture carefully: the rewrite has just shown you what a model thinks your sentence implies. Often that is the fastest way to discover that the phrase you thought was specific was not. Then take the parts you liked, write them into your own prompt in your own words, and turn the switch off.

From that point on you are tuning, and tuning wants one variable moving at a time. Keep the same seed of an idea, change one clause, render again at 40 coins, compare. That loop is slower to describe than to do, and it is the whole skill.

What it does not do

It does not make a refused prompt work. The rewrite deals in detail and structure, not permission, so the limits sit exactly where they did. It does not improve resolution either, which is the separate 60-coin choice between Standard and High. And it is not a quality setting in disguise: a rewritten bad idea is still a bad idea, rendered more confidently.

The honest summary is small. If you wrote a sentence, let it be rewritten. If you wrote a paragraph, you have already done the work, and paying to have it paraphrased is a strange way to spend a quarter more per picture.