Prompt control: the wording you submit is the wording that renders

Why a refusal elsewhere tells you so little

A prompt filter reads your text before anything is rendered and compares it against a list of terms and patterns. When something matches, you get a refusal, and the message is almost always the same regardless of what triggered it. That is why the experience feels arbitrary. The filter is reacting to wording, so a scene that passes when described one way is stopped when described another, and you are left rewriting blind against a rule you cannot see.

The blindness is the expensive part, not the refusal. You cannot tell whether the problem was a noun, a verb, an adjective it read as an intent, or a combination that means nothing on its own. So you try synonyms. The list you are guessing against was tuned for a mass consumer product and for the complaints its operator expects, which has very little to do with what you were asking for.

Models

Vellria Flash

The fastest and cheapest image model. For drafts, variations and quick tests.

Text to image1 credit

Seedream 5 Pro

The flagship image model, focused on detail and composition. 2K output only.

Text to image9 credits

Seedream 5 Pro Edit

The flagship editing model. Takes up to ten references and preserves identity and lighting.

Image to image9 credits

Qwen 3 Pro

Strong on long, complex prompts; stays faithful to scene composition and text.

Text to image6 credits

Qwen 3 Pro Edit

Edits using the images you provide as references, preserving character and scene.

Image to image6 credits

Seedream 5 Lite

Fast, cheap Seedream. Reaches 4K and is good at typography and poster work.

Text to image6 credits

Seedream 5 Lite Edit

Takes up to fourteen references and redraws the images you provide.

Image to image6 credits

The second layer, after the render

The other kind of block runs after the image exists. It is produced, a classifier looks at it, and the result is withheld from you. From where you sit the two are indistinguishable: a wait, then a message. They fail for different reasons and neither says which one happened.

This one has a bill attached. The model did the work, the compute was spent, and on a platform that charges per generation you have paid for a picture you were not allowed to see. It also fails on its own terms fairly often, because a classifier judging a whole image has less to go on than you had when you wrote the sentence.

The quieter failure: silent rewriting

A refusal at least announces itself. The harder problem is rewriting: some platforms expand or reword your prompt before it reaches the model, adding detail, tidying grammar, and smoothing whatever they read as risky. An image comes back, so nothing looks broken. It is simply not the image you described, and nothing tells you your text was edited in transit.

This is what makes prompting feel unreliable elsewhere. You change a word, the rewriter reacts to your change in its own way, and you end up debugging a sentence you never wrote. Some models in this catalog carry that rewriter as a switchable vendor setting and it is sent switched off; the rest have no rewriter to switch. Either way the text you submit is the text that is used.

That is a property of the request, not a preference stored on your account. It holds the same whether the run came from a model page or from your own code, because both build the same body and neither adds a layer the other does not have.

When your wording is the only variable

With nothing rewriting your prompt, the difference between two renders is what you wrote. That makes prompting more precise work rather than less. Describe where the subject sits in the frame, what the light is doing, what the lens is like. A stack of adjectives hands the composition back to the model, which is exactly the ambiguity you were trying to remove.

One input is not yours to set, and it is worth knowing which: the request carries no seed field, so the run-to-run variation underneath your wording is not something you dial in. Read a change across a few runs rather than off a single pair, and change one thing at a time so that what moved between them is the edit you made. Against a rewriter the same experiment tells you nothing, because two variables moved and only one of them was yours.

This is also why a prompt worth keeping is worth saving verbatim. Nothing is inserted between the words and the render, so the sentence that worked is a record rather than an approximation. It carries sideways as well: every image endpoint in the catalog reads the same prompt field, so wording tuned on a cheap model runs on a heavier one without translation, and what differs between them is the render rather than your interpretation of some intermediate layer's habits.

What still gets refused, and why that is not a prompt filter

Fidelity describes the prompt, not the policy. A subject filter guesses at intent from your phrasing and gets it wrong in both directions, blocking ordinary work and missing what it was built for. The refusals we apply are a category rather than a guess: anyone who is not of age, and real, recognisable people used without consent.

The difference matters in practice because a category does not need to read your tone. Nothing has to infer whether you meant something innocently, and nothing loosens or tightens with usage. What that category does not cover is the provider that runs the model, which keeps a filter of its own on the requests it receives, judges them on its terms rather than on the ones published here, and can stop a run once it has started. A refusal from there is recorded as its own kind of failure and the credits taken at the start are returned automatically, so what it costs you is a re-run rather than a balance. The content rules set out the full policy, what happens to an account that crosses it, and who can see the work you generate.

Frequently asked questions