scenario-luma-video

Scenario Luma Video

Overview

Luma's video line on Scenario is four sibling models, one per mode: Ray 3.2 generates from text or frame anchors, Ray 3.2 Edit restyles footage under structural controls, Ray 3.2 Reframe outpaints to a new aspect ratio, and Modify Video makes budget prompt edits. Route by member before tuning anything; Luma's image models belong to the scenario-luma-image skill. Discover the live set with search and treat model_schema_get as the contract.

Connection and the core loop: see the scenario skill in this repo; 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

Each mode is its own model (names from the live schemas, caps as of authoring time):

Member	Inputs	Behavior
Ray 3.2	prompt (+ startFrame, endFrame)	text or image to video; endFrame is only valid alongside startFrame
Ray 3.2 Edit	video + prompt	restyles while keeping motion and timing; editStrength plus controls
Ray 3.2 Reframe	video + prompt + aspectRatio	fits a new ratio by outpainting, never cropping; prompt fills the canvas
Modify Video	video (+ prompt, firstFrame)	budget prompt edits via a mode ladder; source up to 30 seconds, 100 MB

On Ray 3.2 the options veto one another. At authoring time duration was a unit string, "5s" or "10s", never a number; "10s" refused loop, hdr, startFrame, and endFrame. hdr required 720p or 1080p from the 540p, 720p, 1080p resolution list; six aspectRatio values ran 9:16 through 21:9. Edit and Reframe share that resolution list; Modify Video has no resolution parameter. duration, resolution, and hdr each move the price (1080p ran several times 720p's cost and wait), so dry_run the option set before a batch.

Two editors, one strength ladder

Ray 3.2 Edit and Modify Video restyle through the same nine-value ladder, adhere_1 through reimagine_3 (adhere stays close, flex restyles but keeps recognizable elements, reimagine transforms), and share little else. Edit names the ladder editStrength; Modify names it mode. Edit takes a guide image as startFrame; Modify's firstFrame is an edited copy of the source's own first frame, not an arbitrary style image. Edit alone offers keyframes (up to 64 image-and-index anchors, mutually exclusive with the single startFrame guide), face, pose, depth, normals, and trajectory conditioning toggles, resolution choice, and hdr; or pass autoControls: true instead of a manual editStrength. The price gap ran near two orders of magnitude at authoring time, so dry_run the same clip on both and pay for Edit's controls only when the edit needs them.

Prompt motion, not adjectives

Ray rewards natural-language prompts of roughly 50 to 300 words built around motion: a subject mid-action, one named camera move (slow push-in, handheld follow), and one physical consequence of the action (droplets scattering, dust rising). Concrete lighting language lands directly. With loop, hold motion intensity constant (steady rain, drifting steam) so the cycle closes cleanly. On Reframe, prompt only the edges: describe what the new canvas should contain and leave the original subject alone.

Worked example: a 16:9 hero clip recut for 9:16

1. search with target="models", query="luma", public=true. Prefer the newest non-deprecated generation, e.g. model_luma-ray-3-2 and its Edit and Reframe siblings (live hits at authoring time: re-discover each session).
2. model_schema_get with the generator id: options and their vetoes before anything else.
3. upload_asset the product still (see the scenario skill) to get an asset id.
4. model_run with that model_id, dry_run=true, and the exact parameters={"prompt": "A crystal perfume bottle catches soft window light as a single drop arcs off the stopper, slow push-in, product commercial style.", "startFrame": "asset_a", "duration": "5s", "resolution": "1080p", "aspectRatio": "16:9"}; re-estimate after any option change.
5. Repeat model_run with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.
6. model_schema_get the Reframe sibling, then model_run with parameters={"video": "<generated asset id>", "aspectRatio": "9:16", "prompt": "continue the marble counter downward and the softly lit wall upward", "resolution": "1080p"}, same wait discipline.
7. asset_display both and inspect the outpainted edges before delivery.

Common mistakes

• duration: 5 or "5": Ray 3.2 takes the unit string, "5s" or "10s".
• A 10 second loop, HDR grade, or frame anchor: "10s" excludes all three; drop to "5s".
• Ray parameter names on Modify Video: it takes firstFrame and mode, not startFrame and editStrength.
• editStrength with autoControls, or startFrame with keyframes: each pair is mutually exclusive on Edit.
• Prompting the whole scene on Reframe: the prompt describes only the added canvas; the source subject stays.
• Vibrant, whimsical, or hyper-realistic in a Ray prompt: they degrade quality in this family.