Skip to main content
Models that take references (image-to-video, first and last frames, motion control, lip sync, voice reference) take asset ids. The upload flow gives you one for a file on your side, and every generation output already has one, so chaining generations needs no upload at all.
1

Register the file

The response has asset_id (what you will pass to a generation), upload_url (where to PUT the bytes) and required_headers (send them exactly). Send the file’s exact byte count in size_bytes; a smaller, empty or larger PUT is rejected. For video and audio, send duration_seconds too.
2

PUT the file

Send the required_headers from the previous response exactly as given; the declared size is signed as both the minimum and maximum. Successful uploads therefore match the size registered for the asset and reserved against your storage quota. Upload URLs expire after ten minutes. Register the file again if you miss it.
3

Use the asset id in a generation

GET /v1/models/{id} names the reference inputs each model and mode accept (first_frame, reference_images, motion_video, …) and which are required.

Accepted files

Images (PNG, JPEG, WebP, GIF, AVIF) up to 20 MiB, video (MP4, MOV, WebM) up to 1 GiB, and audio (MP3, WAV, M4A) up to 100 MiB. MIME aliases such as audio/m4a, audio/x-m4a and audio/x-wav are normalized; use the returned Content-Type header for the PUT. Audio WebM is not supported by this upload endpoint. Per-model limits on count, duration and frame rate come from GET /v1/models/{id}; a request that breaks them is rejected before any credit is reserved. An asset id must belong to your organization and be visible to the key’s creator, the same rule the app’s Assets library applies. Uploaded files appear in that library like any other upload.