Generate a video from 1-5 reference images and a text prompt (references-to-video). Unlike createVideo, which animates a single source image, this composes a new scene that borrows characters, objects, and style from the reference images. Each image can be a URL or base64. Griffin (default, 480p) and Griffin HD (720p) generate a soundtrack; note the per-second rate here is higher than createVideo's for the same model. The job result is the video URL and its actual duration in seconds. Choose the output shape with `aspect_ratio` ("default" lets the model decide). The chosen `model` and `duration` must be compatible (incompatible combinations return HTTP 400). Credits are held when the job is accepted; the final charge is max(rate × produced seconds, the model's minimum charge), never more than for the duration you requested, and the difference (or everything, if the job fails or is cancelled) is refunded. Pass an optional `request_id` to tag the result so you can locate it later via listGenerations (type video). Related tools: `createVideo` for image-to-video, `editVideo` to modify a generated video. Requires an API key (user scope). The call returns 202 with a job id - poll getApiJob (pass wait 30 to long-poll) honouring poll_after_ms, until status is succeeded; its result field is exactly the response documented for this operation. Synchronous responses are deprecated but still supported: pass `async: false` to block until the result is ready and receive it as the response body. Synchronous calls run on the same queue: the response carries an X-Ludo-Job-Id header, and a job still running after 15 minutes comes back as 202 with the job instead of an error. Each account may have up to 50 generations queued or running at once via the API; a request beyond that returns 429 (code PENDING_JOBS_LIMIT).