Core concepts

Video generation

Same gateway auth as chat. Video jobs can take several minutes, use a long timeout.

POSThttps://api.rodiumai.io/v1/videos/generations

Call POST /v1/videos/generations with a Veo (google/veo-3.1, -fast, -lite) or Sora (openai/sora-2) model id. Text-to-video and image-to-video are supported; Veo also accepts a last_frame. The call is synchronous: the response arrives when the clip is ready.

Long-running jobs

The request stays open until the video is ready, up to about 10 minutes; there is no job id to poll. Set a client timeout of 10+ minutes and read data[0].b64_json (or url).

When to use

Product motion: animate a logo or packaging still.
Short ads and social clips from a text storyboard.
Controlled transitions between two keyframes (last_frame).

Recipes

Text → video

Send model + prompt + duration_seconds. No image field.

Note: Use cinematic prompts; keep duration within the model’s limits.

Image → video

Pass image as { b64_json }, a data: URL, or gs://. HTTP(S) image URLs are rejected by the gateway.

Note: Encode a local PNG/JPEG to base64; do not fetch remote URLs server-side in the request body.

last_frame transition

Provide image (start) and last_frame (end) with the same encodings to morph between two frames.

Note: Both frames should share aspect ratio when possible.

Text → video examples

…

Image → video

…

Request parameters

Parameter	Type	Required	Description
model	string	Required	Video model id (e.g. google/veo-3.1, google/veo-3.1-fast, openai/sora-2).
prompt	string	Required	Text description of the video to generate or animate.
duration_seconds	number	Optional	Clip length in seconds (alias: seconds). Default 8. Sora snaps to 4, 8 or 12. Billed per second of the returned clip.
aspect_ratio	string	Optional	Aspect ratio, e.g. "16:9" or "9:16". On Sora it selects the size when size is absent.
size	string	Optional	Sora: output size "1280x720" (default) or "720x1280".
resolution	string	Optional	Veo: output resolution when the model supports it (e.g. "720p", "1080p").
resize_mode	string	Optional	Veo image-to-video: how the first frame is fitted (e.g. "pad" or "crop").
image	object \| string	Optional	First frame (Veo) or reference image (Sora, resized to the output size). Accepts { b64_json }, a data URL, raw base64 or a gs:// URI; HTTP(S) URLs are rejected. Aliases: input_image, image_url.
last_frame	object \| string	Optional	Veo only: optional end frame (same encodings as image) for start → end transitions.

API reference: videos

A synchronous call

POST /v1/videos/generations keeps the connection open until the clip is ready, usually one to several minutes and up to about 10. There is no job id and nothing to poll: set a client timeout of at least 10 minutes.
Veo models (google/veo-3.1, google/veo-3.1-fast, google/veo-3.1-lite) accept aspect_ratio, resolution, resize_mode, a first frame (image) and a last_frame.
Sora (openai/sora-2) accepts size (or aspect_ratio), duration_seconds snapped to 4, 8 or 12, and an optional reference image resized to the output size.
Authenticate with Authorization: Bearer (no x-api-key). Keep the request body under 10 MiB.

Response

The body has created, model and data[0] with b64_json (MP4, with mime_type) and/or url, plus duration_seconds when the provider reports it. Prefer b64_json when present.

Billing is per second: the hold uses duration_seconds (8 when omitted) and you pay the duration the provider returns. A generation that fails or is blocked by safety filters is not billed.