Magic prompts

Write once.Three prompts.

Set magic_prompt to true. Mynth rewrites it for the model that will run, and the task returns the text that ran.

Get an API key Enhance prompts

You wrote

pls make a product foto of our candle. cozy, wooden table, maybe some plants??

The model reads

Seedream Pro
bytedance/seedream-pro
Rewritten
Product photograph of a candle on a wooden table, with a few plants beside it. Soft warm light.
Model-tuned. Sentences, and the light is named.
Bismuth Illustrious Mix
john6666/bismuth-illustrious-mix
Rewritten
masterpiece, best quality, very aesthetic, candle, wooden table, plants, warm lighting, product photo
Model-tuned. The same note, as tags.
FLUX.2 Pro
black-forest-labs/flux.2-pro
Rewritten
A product photo of a candle on a wooden table, with a few plants beside it, in warm light.
No style card. The general rewrite is a short sentence.

Examples, for one candle. The live rewrite varies. The task returns the text that ran.

The request

One field on the generate body

Turn it on and the model sees the rewrite. Leave it off and the model sees what you sent.

magic_prompt: Set true on image.generate. Omit it, and the model receives your string unchanged. The result then has no magic_prompt.
request.prompt: Your original stays on the task. The rewrite does not replace it.
result.magic_prompt: positive is the text that ran. negative is there only when the request included a negative_prompt, and that string is rewritten too.

Enhance prompts is the call, the CLI flag, and the result shape. Prompts and images covers what the rewrite reads, and what the provider receives.

The rewrite

It keeps what you meant

The pass stays close to your words. A looser, more creative rewrite is in development.

Wrapper phrases: Phrases such as "please generate" and "based on the description below" are removed.
What stays: Subject, counts, names, and any text meant to appear in the image. Quoted text stays verbatim.
What is not added: No new objects, branding, or a direction you did not ask for. Detail is added only when the request strongly implies it.
Already in shape: A prompt already written the way that model wants stays close to what you sent. The reads marked Kept above are that case.

Per model

Some models carry their own instructions

Every image model is rewritten. Some also include instructions for that model, and the rewrite follows them.

Those instructions say how the model wants to be prompted, and they can add cues that raise quality. Some include example prompts. A model without instructions gets the general rewrite, a short run of sentences.

Price

The estimate includes it

Magic Prompt is a line on the generation estimate. Today that line is $0.

$0 today: The generation estimate includes Magic Prompt, so a later price shows before you send. Estimate cost is the request.
Before the image: The rewrite is one text-model pass, and it finishes before generation starts.
Video: video.generate does not take magic_prompt.

A weak rewrite on a specific model is worth saying on Discord (opens in a new tab). A style card is how that model gets its own instructions.