Gemini Omni Flash (Google DeepMind)

Gemini Omni is Google DeepMind's video generation and editing model family. The OpenMontage direct adapter uses the generally available Gemini Omni 1.1 Flash model (gemini-omni-1.1-flash) through the Gemini Interactions API. Its implemented baseline generates 3–10 second clips at 720p/24fps with synthesized audio. Its differentiator in the OpenMontage fleet is stateful conversational editing: each stored generation returns an interaction_id, and a follow-up call with previous_interaction_id edits that video in place — no other wrapped provider can refine a clip without regenerating it.

OpenMontage wraps it as gemini_omni_video (native Gemini API, no gateway). It shares GOOGLE_API_KEY/GEMINI_API_KEY with google_imagen and google_tts — one key, three capabilities. Paid tier only: ~$0.10 per second of output video (exactly 5,792 output tokens/sec at $17.50/1M, before input and any text/thought output).

Other documented routes are available when the direct Google key is not the chosen provider:

Use the direct gemini_omni_video route for stateful conversational editing. Gateway routes return ordinary provider tasks and cannot preserve Google's previous_interaction_id workflow. The fal edit endpoint can still be iterated by feeding each output video URL into the next edit call.

When to pick it (and when not)

Route through video_selector for generation operations. Editing (edit_video) is a direct-tool operation — call gemini_omni_video from the registry, because the multi-turn interaction state lives outside the selector's model.

Billing and duration hints

• duration is only a planning/prompt hint for the caller. The adapter does not send it to Google; express timing in the prompt or timecodes, and let the model choose the actual length. It does not cap billing.
• ToolResult.cost_usd uses valid response usage when available: input tokens are $1.50/1M for every modality, output text is $9.00/1M, and output video is $17.50/1M. Unknown/unclassified output and separately reported thought tokens use the higher $17.50/1M rate so the result cannot underreport.
• Without valid usage, the fallback assumes the maximum 10-second output: 10 × 5,792 × $17.50/1M = $1.0136, plus a conservative upper estimate for prompt and available media input tokens. A short duration hint never reduces this fallback.

Generation prompting

Describe scene + camera + lighting + motion + audio. Official example:

Continuous, unbroken handheld shot of a fluffy tabby cat sitting on a sunny windowsill, looking out into a leafy garden. The cat's tail twitches slowly, and its ears rotate slightly toward ambient noises. Sunbeams illuminate dust motes in the air.

• Force a single shot explicitly: "In a single continuous shot," / "No scene cuts." Otherwise the model may cut between scenes.
• Negatives go in prose — there is no negative_prompt parameter: "No dialogue," "No extra sound effects."
• No sampler controls: system instructions, temperature, top_p, and seeds are all unsupported. The prompt is the only lever.
• Meta-prompt for quality: "Consider micro-detail, expression and timing to create a very rich, detailed but entirely natural scene."

Timecode syntax

Schedule beats with bracketed ranges or natural language — this maps directly onto OpenMontage scene-plan timings:

[0-3s] A person is walking [3-6s] They stop and turn around

"After 3 seconds, a woman enters the scene." / "At 5s the chorus starts in the background audio."

Audio and on-screen text

Audio is synthesized automatically; direct it in the prompt: "Include calm background music," "The audio is a low tinny radio broadcast in the background." Rendered text works and can be timed:

One word on the screen at a time: 'did, you, know, that, Omni, can, do, awesome, text?' Each word appears for 1s.

Reference images (<FIRST_FRAME> / <IMAGE_REF_N> tags)

Pass local images via reference_image_paths (they are sent in order), then bind them to roles inside the prompt with tags. <IMAGE_REF_N> indexes from 0 in the order supplied:

in the style of <IMAGE_REF_0> a woman <IMAGE_REF_1> is walking

[0-3s] A studio fashion sequence. Starting with woman <IMAGE_REF_0>, she is
holding <IMAGE_REF_1> [3-6s] Then we see the man <IMAGE_REF_2> holding <IMAGE_REF_3>

• <FIRST_FRAME> makes an image the opening frame: <FIRST_FRAME> a woman is walking.
• Use high-resolution images; describe the intended motion specifically rather than "make it move."
• Say what each image is (product / character / style / background reference) — the model decides usage from context.

Conversational editing (the differentiator)

Editing prompts are the opposite of generation prompts: short and surgical. Overly descriptive edit prompts cause unintended changes.

1. Generate the base clip (subject + scene + motion). The tool returns interaction_id in its result data.
2. Pass it back as previous_interaction_id with operation="edit_video" and describe only the delta.
3. Append "Keep everything else the same." to pin unmentioned elements.
4. Refine in layers — one turn for lighting, one for camera, one for action, one for audio.

Official good/bad pairs:

Other working edit prompts: "Make this video anime" / "Put a fashionable hat on this person" / "Change the lighting to be more dramatic" / "Change the text on the sign to say 'Omni Flash'".

Gotcha — store: editing via previous_interaction_id only works if the prior call kept the interaction server-side (store defaults to true in gemini_omni_video). Set store=false only for one-shot generations you will never edit.

Download recovery: after an accepted paid generation, a Files API poll or download can still fail. In that case the tool returns output_uri, the usage-based charged cost when valid usage was returned (otherwise the conservative fallback), and a resume_download payload. Pass that payload back to gemini_omni_video to retry only the existing file's GET/poll flow; it must not submit another paid generation.

Editing uploaded videos: pass input_video_path instead of previous_interaction_id; the tool uploads it via the Files API. Unavailable in the EEA, Switzerland, and the UK (editing generated videos works everywhere).

Current adapter limitations (stable model)

• Output: 3-10s, 720p, 24fps, MP4 with audio; aspect ratio 16:9 or 9:16. All output carries an invisible SynthID watermark.
• No seed, negative prompt, temperature, top_p, or system instructions.
• The stable Google model supports video extension, first/last-frame interpolation, and additional resolutions, but this OpenMontage adapter has not implemented those controls. Do not advertise them as working direct-tool features. No voice editing.
• Audio reference inputs unsupported. Video references ≤3s are accepted by the schema but not processed correctly — don't rely on them.
• Multi-video prompting unsupported; may degrade output.
• English fully supported; other languages untested.
• Images of minors (EEA/CH/UK) and certain recognizable people are blocked for upload/editing.

Sources

• Generation & editing guide: https://ai.google.dev/gemini-api/docs/omni
• Model card: https://ai.google.dev/gemini-api/docs/models/gemini-omni-flash
• Pricing: https://ai.google.dev/gemini-api/docs/pricing
• Announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/