Generate

One call. What comes back depends on the modality, not on how you asked.

ts
const handle = await sequel.run({
  model: 'black-forest-labs/flux-1-schnell',
  params: { prompt: 'a ceramic mug on seamless white', steps: 8 },
});

handle.task.assets?.[0]?.url;

params is the model’s own input

Most models take prompt. Text-to-speech takes text. There is no normalised field, so prompt exists as a shorthand that folds into params rather than replacing it.

ts
await sequel.run({ model: 'openai/gpt-image-2.5-flare', prompt: 'a mug' });
await sequel.run({ model: 'deepgram/aura-1', params: { text: 'Your order is on its way' } });

Several at once

They go in parallel. One failure does not take down the others.

ts
const handles = await sequel.runMany([
  { model: 'flux', prompt: 'a mug' },
  { model: 'flux', prompt: 'a wallet' },
]);

Reference images

Upload, then pass the URL. The server fetches and encodes it for providers that want base64, so you never deal with that yourself.

ts
const url = await sequel.upload(file);

await sequel.run({
  model: 'openai/gpt-image-2.5-flare',
  params: { prompt: 'turn this into a silk scarf', images: [url] },
});

Check the schema first — the field differs by model (images, image_url, first_frame_image), and a text-to-image model ignores a reference silently.

Writing a good prompt

  • Prose, not comma-separated tags. These encoders are trained on sentences.
  • Concrete over decorative. Materials, lens and lighting move the result; “8K, masterpiece” does nothing.
  • Say what should be in frame, never what should not. “No people” tends to summon people.