When the AI rewrites your prompt before it draws

You asked for one thing and the image came back with extras: a sunset you never mentioned, a landmark you didn't name, a mood you didn't choose. It's tempting to blame the model for misunderstanding you. Often it understood you fine, and the extras came from a rewrite of your prompt that happened before the model ever saw it.

The extra details usually aren't a misunderstanding

Microsoft's documentation for DALL·E 3 on Azure OpenAI shows the mechanism with a plain example. The prompt sent in was "Watercolor painting of the Seattle skyline". The prompt the image model actually received — the revised prompt — began "A soft and vivid watercolor painting capturing the scenic beauty of the Seattle skyline," and went on to describe "a setting sun casting warm hues over the sprawling cityscape, with the Space Needle prominently standing tall against the sky," the water of Puget Sound and the snow-covered peak of Mount Rainier in the distance.

None of that contradicts the request. All of it was added. If the finished image had a sunset and a mountain in it, the honest answer to "where did those come from?" would be: a sentence the person never wrote.

That model has since been retired — Microsoft's page notes that the DALL-E model dall-e-3 was retired on March 4, 2026 — but the documentation is still one of the clearest public descriptions of how this kind of rewriting works.

Prompt rewriting is a documented feature

What it adds, and why

Microsoft calls it prompt transformation and describes it in a few sentences: "Prompt transformation is a process included in the DALL-E 3 models that applies a safety and quality system message to your original prompt. It uses a large language model (LLM) to add a message before sending your prompt to the image generation model." The point, in Microsoft's words, is that "This system message enriches your original prompt with the goal of generating more diverse and higher-quality images while maintaining intent."

Two things in that description matter to you. The rewriting is done by a language model, before the image model sees anything. And its job is to keep your intent while adding detail, which is exactly why the additions tend to feel plausible rather than wrong.

OpenAI's documentation describes the same pattern for its own image generation tool. In its Responses API, the main model "will automatically revise your prompt for improved performance." The guide's sample request asks for a gray tabby cat hugging an otter with an orange scarf, and the sample revised prompt on the same page reads: "A gray tabby cat hugging an otter. The otter is wearing an orange scarf. Both animals are cute and friendly, depicted in a warm, heartwarming style." The request never chose a style. The rewrite did.

Some tools let you turn it off, some don't

Microsoft's page left no room for doubt about its own service: "You can't disable or customize prompt transformation for DALL·E 3 deployments in Azure OpenAI. The revised prompt is always generated automatically." Other tools put a switch in front of you — Vipix has one under the prompt box, and what it does is explained below. Before you assume either, check what the tool you're using actually offers.

How to tell whether your prompt was rewritten

Both companies' developer documentation shows where the rewritten version comes back: a field called revised_prompt. Microsoft: "You can see your revised or transformed prompt in the API response object, in the revised_prompt field." OpenAI: "You can access the revised prompt in the revised_prompt field of the image generation call."

That's the developer's view. In an app built on top of a service like these, whether that text ever reaches your screen is the app's decision. So the practical rule is simple. If your tool shows you the prompt it actually used, read it — it's the most direct explanation you'll get for anything unexpected in the picture. If it doesn't show one, don't conclude that nothing changed. Read your own prompt back instead and look for what you left open.

Short prompts leave the most room

Here's the idea that makes all of this predictable: whatever you leave blank, something fills in. An image can't leave the lighting, the background or the camera distance undecided. If your prompt doesn't decide those things, something else will — the rewriting step if there is one, the image model if there isn't.

So the shorter your prompt, the bigger the share of the picture that gets decided for you. Write four words and most of the image is someone else's choices. Describe the things you actually care about and that share shrinks. The Seattle example shows it well: the request decided "watercolor" and "Seattle skyline," and everything else — the time of day, which landmarks, what sits in the distance — was decided downstream.

It also explains why two people can have opposite experiences with the same tool. The one who writes full prompts finds it obedient; the one who writes three words finds it pushy. Same mechanism. The difference is how much open ground each of them left.

What to actually put in a prompt is its own subject, and our AI photo editing prompt library goes through it. If you're generating images to sell products, whether AI images belong in a product listing is a separate question worth settling first.

What Vipix does with an English prompt

Vipix has a switch of its own, and it's only fair to say plainly what it does — because for most prompts written in English, the answer is nothing at all.

The switch sits under the prompt box in Generate images and is on by default. Its label reads "Rewrite Vietnamese prompts in English", and the note under it says that only prompts with Vietnamese-style accented letters are rewritten. It was built for prompts written in Vietnamese: when a prompt contains Vietnamese letters, it's translated into English and enriched before the image is made, and if you're logged in, the editor shows a line beginning ✨ Prompt (EN): with the wording that was actually used.

A prompt in plain English skips that step entirely. It goes to the image model exactly as you typed it — no translation, no rewriting — so there's no ✨ Prompt (EN) line to read either, because there's nothing different to show. Type a stack of old books by a window with the switch on or off and the same words go through.

There is one exception, measured in the code as it stands today. The check for Vietnamese counts accented letters that Vietnamese shares with other languages, so an English prompt containing words like café, décor, rosé, résumé or à la carte is treated as Vietnamese and goes through the translate-and-enrich step, which means it can come back longer than you wrote it. Words like jalapeño, piñata, façade or naïve don't trigger it. If your prompt has accented words like those and you want your exact wording, switch the option off.

Either way, Vipix checks both what you typed and any rewritten version. The same switch, with the same behavior, sits under the prompt box in Generate photorealistic images.

When to keep your exact words

Two situations call for turning any rewriting off, on any tool that lets you.

The first is when you've already written a careful prompt and weighed every word. A rewriting step at that point only produces a sentence you didn't need anyone to write for you.

The second is when you're testing changes. If you want to know what one edit to your prompt does, you want the runs to differ only where you changed something. Turning rewriting off removes one moving part — a freshly rewritten sentence each time — though each run is still a new variation, so judge the trend over a few runs rather than a single pair.

On Vipix, with a prompt in plain English, the switch makes no difference either way; it only matters when your prompt contains accented words like the ones above.

Frequently asked questions

Why does the AI add things I didn't ask for?

Many AI image tools rewrite your prompt before the image is made, adding detail to improve quality while trying to keep your intent. Microsoft's documentation for DALL·E 3 shows a request for a watercolor of the Seattle skyline turned into a description with a sunset, the Space Needle and Mount Rainier. Anything your prompt leaves open gets filled in.

Can I see the rewritten prompt?

Sometimes. Microsoft's and OpenAI's developer documentation both return it in a field called revised_prompt, but an app may or may not show it to you. On Vipix, a prompt in plain English isn't rewritten, so there is nothing extra to see; prompts that do go through the rewriting step show the wording that was used when you're logged in.

Can I turn prompt rewriting off?

It depends on the tool. Microsoft's documentation said it couldn't be turned off for DALL·E 3 on Azure OpenAI. On Vipix there is a switch under the prompt box, and it only acts on prompts that contain Vietnamese-style accented letters.

Does Vipix rewrite my English prompt?

Not if it's plain English: the prompt goes to the image model exactly as you typed it. As the code stands today, a prompt containing accented words such as café is treated as Vietnamese and rewritten, so switch the option off when you need your exact wording.

Will a longer prompt stop the rewriting?

No. A longer prompt doesn't switch off another tool's rewriting step, but it leaves less for that step to fill in. It narrows the gaps, but it can't make the image match your words exactly.

Write it, read it, then generate

Write the prompt, read it back for the things you left open, then generate with Generate images. When a generation uses credits, the editor shows its cost before you start. And once the image arrives, checking an AI portrait before you post it picks up where this article stops.