AI Character Reference Bank

Give it a one-line character description and it generates a matching set of reference images (portrait, three-quarter, full body, side profile, and expression sheets) plus a reusable "character bible" prompt you can paste into later jobs. It's built for people who need the same character to show up identically across many shots: comics, faceless videos, game concepts, storyboards.

How it works

For each character it writes a character-bible prompt from your description and shared style, then sends that same bible to OpenAI's gpt-image-1 once per requested view, appending the view instruction (front portrait, 3/4 angle, full body, side, or an expression). Reusing one locked description per view is what keeps the face, hair, clothing, and colors consistent across the set.

Input

You bring your own OpenAI key. Nothing is strictly required by the schema, but in practice you need characters and openaiApiKey or the run errors out.

Output

Each character is one dataset record. The PNG files land in the run's key-value store, and the record links to them. Key fields per record: name, slug, description, characterBible (the reusable prompt), and references, an array of { view, imageKey, imageUrl }. If a particular view fails, that entry carries an error instead of an image and the rest still run.

Example

{
  "characters": [
    { "name": "Milo", "description": "a small brave fox cub with a red scarf and bright green eyes, cartoon style" }
  ],
  "views": ["portrait", "threequarter", "fullbody", "happy", "angry"],
  "style": "consistent character design, clean studio background, soft even lighting",
  "aspectRatio": "1:1",
  "quality": "high",
  "openaiApiKey": "sk-..."
}

Pricing

$0.15 per character, pay per result, no subscription. OpenAI image costs are billed separately to your own key.

Notes

The character bible reduces drift but doesn't eliminate it. Across many views you'll still see small inconsistencies, so treat the set as strong references rather than a guarantee, and reuse the characterBible string verbatim in downstream prompts to stay on model.