scenario-runway

Scenario Runway Video

Overview

Runway's family on Scenario splits by direction: Gen4.5 generates new clips from text or a first-frame image; Aleph 2 rewrites existing footage, changing what the prompt names while motion, audio, and structure carry over. Discover both with search; model_schema_get is the contract.

Connection and the core loop: the scenario skill; model-agnostic video work: the scenario-video skill. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Mode follows from the inputs (names and caps from the live schema):

At authoring time Aleph capped sources at 30 seconds and 16 MB, separate limits: high-bitrate clips hit the size cap first. Its median cost ran six times Gen4.5's, a fifteen-fold spread following the source video: dry_run the exact payload first. Median latency neared six minutes, so jobs_wait will time out mid-job; re-call it with pending_job_ids.

Gen4.5 prompts follow the mode

Text mode needs both the visuals and the motion described. A first frame fixes subject, composition, and palette: spend the whole prompt on motion, camera, and atmosphere. Order beats explicitly ("kneels, then rises, camera cranes up") and give multi-beat prompts 8 to 10 seconds of duration. Name camera moves plainly (locked camera, pan, dolly, tracking, orbit, crane), one or two per clip, in present-tense prose.

aspectRatio takes pixel pairs ("1280:720", "960:960"), and text mode honored only the 16:9 and 9:16 pairs at authoring time; get other shapes (1:1, 4:3, 3:4, 21:9) from a first frame with that geometry instead.

Aleph changes what you name

An Aleph prompt states the edit: remove, swap, restyle, relight. Everything unnamed persists, along with motion and audio: Aleph cannot re-choreograph or add camera moves; when motion must change, generate with Gen4.5. To steer the edit, pin up to 5 reference images: keyframes places each image (uri) by seconds or at (a 0 to 1 fraction); promptImage by positionType plus timestampSeconds or positionPercentage. The arrays are alternatives, never combined.

Worked example: swap a product in real footage

1. search with target="models", query="runway", public=true, e.g. model_runway-aleph-2 for editing, model_runway-gen4-5 for generation (live hits at authoring time: re-discover each session).
2. model_schema_get with the Aleph id before anything else.
3. Trim or compress a source over the caps (see scenario-video); upload_asset the clip and product still (see the scenario skill).
4. model_run with dry_run=true and the exact parameters={"prompt": "Replace the soda can with the bottle from the reference image. Keep everything else unchanged.", "video": "asset_clip", "keyframes": [{"uri": "asset_bottle", "at": 0}]}; re-estimate after any trim.
5. Repeat with wait=false, then jobs_wait the returned job id, re-calling with pending_job_ids; a timeout is not a failure and never justifies a second model_run.
6. asset_display the output; check the swap holds in every shot.

Common mistakes

• Sending both keyframes and promptImage: exclusive alternatives; pick one.
• Restating the first frame in an image-mode prompt: prompt motion and camera instead.
• Passing aspectRatio as "16:9" or alongside image: pixel pairs only, and a first frame overrides it.
• Asking Aleph for new motion or camera moves: the source's motion persists by design.