scenario-gemini-omni

Scenario Gemini Omni Video

Overview

Gemini Omni, Google's video family on Scenario, splits its modes across catalog members instead of folding them into one model: a text and first-frame generator (ranked first for text-to-video in public arena voting at authoring time), a reference-to-video member for subject consistency, an edit member that restyles existing footage, and on the newer 1.1 Flash line an extend member that appends footage to a clip. Every member generates audio in the same pass. Pick the line and the member by mode with search, reading ids and descriptions rather than rank (the newer 1.1 members ranked below the original line at authoring time), then treat model_schema_get as the contract. Gemini image models are the scenario-gemini-image skill's domain; Gemini TTS belongs to audio.

Connection and the core loop: see 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

Two lines at authoring time, Flash and 1.1 Flash, one member per mode (names from the live schemas):

At authoring time both generators took duration 3 to 10 seconds (default 8) and aspectRatio 16:9 or 9:16. Edit exposes neither knob: length and shape follow the source clip. Extend inherits the shape too (no aspectRatio) but takes duration 3 to 10 (default 6) as footage appended after the source, so an 8 s clip plus 6 returns about 14 s. The original Flash members render 720p only; the 1.1 Flash line adds resolution on every member: 360p a faster draft tier, 720p the default, 1080p and 4k upscaled. The tier does not move price (at authoring time a generation quoted the same at 360p and 1080p, and so did Extend), so a cheap draft is a shorter duration, not a lower tier. Seed and negative prompt exist nowhere in the family. Cost moves with duration, a first-frame image, Reference-to-Video's references, and Edit's and Extend's source clips; Edit spanned the family's widest cost range at authoring time and its jobs ran about twice as long, so dry_run before editing anything long.

Sound is prompted, not switched

No member has an audio parameter: every clip arrives with generated dialogue, ambience, and effects, and the prompt is the only lever. On the generators, end the prompt with an explicit audio line naming what should be heard ("crackling campfire, distant owls") and put spoken words in quotes; leave it out and the model chooses for you. On Edit, never prompt audio: it regenerates to match the new look on its own.

Identity from images, change from words

Reference-to-Video takes the subject's look from referenceImages, and re-describing it in the prompt fights the images. Call the subject "the character" or "the product" and spend the words on action, setting, camera, and sound; several angles of one subject tighten the hold, distinct subjects share one scene, and with no prompt at all the subjects still appear. There is no first-frame anchor here: when the clip must open on an exact composition, pass that still as image on the base member, which also accepts references alongside it and, on 1.1, a lastFrameImage to interpolate towards (it needs the first frame).

Edit is the inverse: motion, camera path, and timing are locked from the source video, only the look moves. Lead the prompt with the change ("make it winter", "swap the red car for a vintage blue Beetle"), name the target concretely, and never ask for new choreography, cuts, or camera moves: those instructions will not take.

Worked example: a mascot short with a consistent character

1. search with target="models", query="gemini omni", public=true. Pick the member by mode, here model_google-omni-flash-r2v (a live hit at authoring time: re-discover each session).
2. model_schema_get with that id: required fields, caps, and defaults before anything else.
3. upload_asset two or three angles of the mascot (see the scenario skill) to get asset ids.
4. model_run with that model_id, dry_run=true, and parameters={"prompt": "The character skips across a rain-slick plaza, catches a falling leaf, and holds it up in triumph. Overcast soft light, low tracking shot. Audio: light rain, footsteps on wet stone, one bright chirp.", "referenceImages": ["asset_a", "asset_b"], "duration": 8, "aspectRatio": "16:9"}; references and duration move the cost, so re-estimate after changing either.
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. asset_display the output and review it with sound: the audio is part of the deliverable.

Common mistakes

• Describing the reference subject's appearance in the prompt: identity comes from the images; write action and setting around "the character".
• Asking Edit for new motion, cuts, or re-timing: they are locked to the source; prompt only the look change.
• Passing duration or aspectRatio to Edit, or aspectRatio to Extend: not in their schemas; the output follows the source clip.
• Hunting for an audio toggle: none exists; steer sound in the prompt or accept the model's choice.
• Restating a first-frame image as a static scene: describe the motion continuing from it.
• Packing a multi-scene story into one run: 3 to 10 seconds holds one continuous beat; carry a beat past 10 seconds with Extend, or sequence shots with the scenario-video-assembly skill.