Gemini Omni Flash References: Image and Video Roles, Resolution, and Response Delivery

Model references should be typed, not guessed

Google’s Gemini Omni Flash guide distinguishes a first frame, a last frame, a style or subject image reference, and short video references. These roles affect the intended operation. Media tags such as FIRST_FRAME, LAST_FRAME, IMAGE_REF and VIDEO_REF can make that intent explicit, while the task field can select text-to-video, image-to-video, reference-to-video, edit or extend. Treat these as guidance rules belonging to Omni 1.1, not generic prompt syntax that every vendor understands.

Reference footage is not a soundtrack source

Google states that video references are best used for likenesses and that audio within a video reference is ignored. The documentation limits video references to three clips of up to three seconds each. A director cannot assume that supplying an actor’s dialogue recording inside a visual reference will make Omni reproduce the speech. Log consent for each referenced likeness and route voice tasks to an audio-specific approved workflow when timing and rights are critical.

Native resolution versus enlarged output

The Omni documentation lists 360p, 720p default, 1080p upscaled and 4K upscaled output choices. Upscaled means an enlarged or processed output, not evidence that the original generative render contained native 4K scene detail. Editors should record the source generation and target export separately. A marketing comparison based only on pixel count can hide that distinction. Review texture, edge artifacts and typography on the actual delivery display before committing to a projection or print-screen application.

Deliver videos without breaking API response handling

For larger outputs Google recommends URI-based delivery rather than embedding big binary payloads. Its REST response wraps video under interaction steps, while SDK convenience fields may differ. A client that assumes every response contains a particular SDK-only `output_video` path may lose completed jobs even when the generation succeeded. Archive response identifiers, check task completion, validate content type, and download into a managed asset store with checksums before rendering a public webpage.

Scenario: consistent character plus changing set

Suppose a documentary animation needs one recurring historian wearing the same coat in two rooms. Use a rights-cleared character image as an appearance reference and a different frame as the intended shot opening; do not label both inputs merely as reference. Prompt the walking motion and camera direction explicitly. If the next shot needs the same voice, do not rely on silent visual-reference clips to determine speech. Inspect face contours, hands, apparel logos and scene transitions independently.

Restrictions and safety

Editing certain recognizable people and image material involving minors may be restricted, particularly in specified regions. The guide also warns that referring to multiple videos is not a general-purpose multi-video reasoning feature. Neither a tag nor a reference grants permission to use someone’s likeness. The model’s content filters and region restrictions are part of production compatibility. Avoid unsafe workarounds and document any rejection without repackaging the footage.

Editorial boundary

This is a versioned API input-and-output guide, not an assertion that Omni achieves perfect identity consistency. Limits, prompt roles and output resolution derive from Google documentation checked October 25, 2026. Pricing, actual latency and plan quotas remain null pending AI10 and project-level verification.

Sources & Further Reading

• Google — Gemini Omni Flash video workflows (first-party source; checked 2026-10-25)