Runway Video HTTP

A reusable Gen-4.5 video-task submission graph built entirely from generic Subfork nodes. It contains no environment-specific graph IDs, composite references, credentials, or generated media. Import it independently, or publish it as a composite before building higher-level video workflows.

Inputs and outputs

Input	Type	Meaning
`prompt`	`str`	Required description of the video
`image_url`	`url`	Optional public HTTPS first-frame image; absent or empty selects text-to-video
`duration`	`int`	Defaults to 5 seconds; provider supports integers from 2–10
`ratio`	`str`	Defaults to `1280:720`; `720:1280` also works in both modes

Outputs are job (resumable job reference), task_url (typed polling URL), task_id, and submission_response (raw provider submission JSON). This graph submits a task; it does not wait for or return the finished video.

Configure RUNWAYML_API_SECRET in Graph Settings → Execution Profile on the executing graph. When used as a composite, the executing parent owns the secret; publication does not distribute credentials. Supply graph inputs through the Python/API execution interface, or connect inputs on the published composite.

HTTP behavior and cost

The graph sends one paid POST to the fixed origin https://api.dev.runwayml.com:

/v1/text_to_video when image_url is absent or empty;
/v1/image_to_video when an image is supplied.

Both requests use model gen4.5, bearer authentication, and X-Runway-Version: 2024-11-06. The endpoint choice is restricted to these two paths; callers cannot replace the credential-bearing origin. The job node uses runtime HTTP defaults of a 30-second timeout and 1,000,000 response bytes. Image URLs must be reachable by Runway and return an accepted image without redirects. Localhost and local file paths are not suitable. The provider validates prompt lengths, image constraints, duration, aspect ratio, and account limits.

Every execution can consume provider credits and Subfork quota. If a submission has an uncertain outcome, inspect the existing execution/task before trying again; automatic resubmission can create another paid generation.

Compose a complete video graph

After importing, validate and publish the adapter as v0.2.1 (or a deliberately chosen new version for changed content). Obtain the generated composite node manifest from the target instance. Build a parent with:

Editable prompt and ratio controls connected to the composite; optionally bind its image_url and duration inputs.
job connected to HTTP Poll Job’s remote_job input.
task_url connected to the poll node’s exposed status_url input.
Poll media connected to Media Response or a Video Preview panel.

Use these poll parameters:

{
  "headers": {
    "Authorization": "Bearer ",
    "X-Runway-Version": "2024-11-06"
  },
  "status_path": "status",
  "complete_values": ["SUCCEEDED"],
  "failed_values": ["FAILED", "CANCELED", "CANCELLED"],
  "media_url_path": "output.0",
  "poll_after_seconds": 10,
  "max_attempts": 180,
  "max_elapsed_seconds": 1800
}

Expose status, raw task response, and video media from the parent. The runtime normalizes the status to lowercase. Poll failures stop the graph. Exhausting polling bounds does not cancel the provider task. Runway’s output URLs expire in 24–48 hours: download videos to lasting storage separately, using an unauthenticated request to the media URL, not the Runway authorization header.

Keep polling at the top level. In the tested development runtime, a composite containing a waiting child lost child resume state and repeated the submission. The submission-only adapter completes once; the parent then resumes just its poll node. Do not nest the complete asynchronous parent inside another composite until that runtime limitation is addressed.

Recreating parents on another instance

A raw parent export containing a development graph ID is not portable. For now:

Import and publish this leaf on the target instance.
Record a tooling-level mapping from logical adapter name to target graph ID and exact published version.
Use the target publication’s manifest to fill the parent’s composite node_id, node_version, and params.composite_graph reference.
Validate the assembled parent before publishing or executing it. Configure target-owned secrets separately.

Pin the reviewed version. Graph IDs differ between environments, but the same reviewed release can retain the label v0.2.1. An existing version with different content must not be silently reused or overwritten. A missing version should produce an installation error, not a fallback to latest.

Publisher-scoped slugs and portable references are proposed features, not current graph fields. They can eventually identify the adapter independently of its local graph ID while retaining exact-version pins. A slug still needs the intended publisher/catalog and the requested release installed on the target. This example deliberately does not introduce unsupported reference fields or ship parent graphs with development IDs.

Validation status

The v0.2.1 source adapter and prompt-only parent were validated and published on development. One five-second prompt-only generation completed on 2026-10-10; the returned MP4 was readable and the user confirmed Video Preview playback. Seven mocked runtime tests covered omitted, empty, and present image inputs, endpoint selection, single submission across polling resumes, failure and cancellation, missing task ID, and polling limits.

This portable export is checked against production node contracts. It has not been freshly imported and executed on production. Live image-to-video generation remains untested. Catalog validation is not evidence of production provider execution. No additional paid generation was performed when adding this example.

The endpoint split follows Runway’s official text-to-video schema and image-to-video schema. See also input requirements and output lifetime. A getting-started example that suggested omitting the image on the image-to-video endpoint did not match the observed API behavior; that request returned HTTP 400.