Skip to main content
A model id has three parts, {brand}/{version}/{task}, and it is also the path of the endpoint that creates a task on that model:

​
Catalogue

Audio: Video Fast 1.0 generates audio with every video, there is no switch. generate_audio marks the endpoints where audio is optional and off by default; — means the task has no audio option. Prices per model, resolution and duration are listed in the developer console. See Credits and pricing for how credits work.

​
Model families

​
Video Fast 1.0

Fast, low-cost generation. A good default for previews, social clips and high-volume use.
  • Tasks: text to video, image to video, reference to video
  • Resolutions: 480p, 768p
  • Duration: 3–15 seconds; reference to video 3–30 seconds
  • Audio: always on. Every Video Fast video comes with generated audio; there is no generate_audio field
  • Aspect ratios: 9:16 (default for text to video), 16:9, 1:1, 3:4, 4:3; image to video defaults to adaptive, which keeps the shape of your image
  • Image input: JPG, PNG or WebP, shorter side at least 300 px, aspect ratio between 0.4 and 2.5, at most 30 MB

​
Wan 3.0 Spark

Native videos up to 30 seconds with richer detail, and synchronized audio on request.
  • Tasks: text to video, image to video, reference to video
  • Resolutions: 480p, 720p, 1080p
  • Duration: 2–30 seconds
  • Audio: optional, generate_audio: true on text and image to video (off by default)
  • Aspect ratios: 16:9 (default for text to video), 9:16, 1:1, 3:4, 4:3; image to video defaults to adaptive
  • Image input: JPG, PNG or WebP, 360–2,000 px per side, at most 20 MB

​
Wan 3.0 Prime Spark

The same Wan 3.0 quality at the full 30 seconds, noticeably faster. Choose it when turnaround time matters.

​
Video Fast Extend 1.1

Continues an existing clip smoothly, with matched audio when generate_audio is true.
  • Task: video extend
  • Resolutions: 480p, 720p, 1080p
  • Duration: adds 3–30 seconds; the clip plus the added seconds may not exceed 30 seconds
  • Clip input: MP4 or MOV, 360–2,000 px per side, 2–30 seconds long, at most 500 MB, with an optional start_time / end_time window

​
Choose a model

Start with short, low-resolution tasks while you tune your prompts, then raise the resolution and duration for the final render.

​
Seed

Every model takes seed, an integer from 0 to 2147483647. Omit it and the model picks a random seed (the seed used is not returned, so keep the one you send). A seed reproduces a result only together with the exact prompt the model saw:
  • prompt_expansion: none — the same seed with the same request reproduces the video: exactly on Video Fast 1.0, and reproducibly but not guaranteed identically on Wan 3.0 Spark / Prime Spark (the model is probabilistic).
  • basic or advanced (the default) — the prompt is rewritten by a language model on every request, so the model sees different text each time and the same seed gives a different video.
Turning expansion off puts the whole result on your prompt: short or vague prompts give noticeably weaker videos, so write a detailed, cinematic prompt yourself (see Write better prompts), or keep expansion on and accept that runs vary. There is no negative_prompt: none of the models takes one, and a request that sends it answers 400 invalid_parameter.

​
Read the catalogue from the API

GET /v1/models lists the enabled models. Each model carries input_schema, the JSON Schema of its request body, which is the authoritative list of fields and allowed values, and x-max-input-bytes, the largest accepted input file.
New models are added without a version change: read GET /v1/models instead of hard-coding the list when you build a model picker.

​
Common request fields

Every create endpoint accepts these optional fields in addition to its own: Text, image and extend endpoints also accept prompt_expansion (none, basic or the default advanced): with basic or advanced the prompt is rewritten into a more detailed description before generation (the rewritten text is not returned). Use none to keep your prompt exactly as written.