多模態模型
Gemini Omni 1.1 Flash Extends Video Using Ten Seconds of Context, but 4K Output Is Upscaled
Google has moved its conversational video model into general availability, adding first-and-last-frame interpolation, scene extensions of up to 40 seconds, and four resolution settings. The new interface provides greater workflow control, but both 1080p and 4K are produced through upscaling, while complex motion and text consistency remain limited.
Google released `gemini-omni-1.1-flash` on August 27, moving the conversational video generation and editing interface previously available in preview into general availability. Through the Interactions API, the model accepts text, images, or videos up to ten seconds long and outputs videos with audio that are three to ten seconds long at 24 FPS. Developers can use `previous_interaction_id` to continue a previous generation or explicitly specify tasks such as `text_to_video`, `edit`, and `extend`.
The most practically significant change is that, when extending a video, the model no longer references only its final second. It now reads the last ten seconds of visuals, motion, characters, and audio. Each operation can add ten seconds, up to a cumulative limit of 40 seconds. The system may modify the final few frames of the input to smooth the transition. Consequently, media pipelines that depend on original-frame hashes, timecodes, or frame-by-frame approvals cannot treat an extended result as content that has simply been appended.
Another new interface allows users to provide a first frame and a last frame, with the model generating the camera movement between them. This is suitable for tracking shots, orbiting shots, and looping transitions. The resolution parameter supports 360p, 720p, 1080p, and 4K. Compared with the default 720p setting, Google says 360p drafts can increase system throughput by up to 60% at roughly one-third of the cost, allowing products to explore compositions in bulk before selecting final outputs. However, the API documentation explicitly identifies 1080p and 4K as upscaled outputs, so they should not be interpreted as containing an equivalent level of natively generated detail.
General availability also starts the countdown to the migration deadline: `gemini-omni-flash-preview` is scheduled to be discontinued on September 30, and applications must update the model name and regression-test prompts, output dimensions, and content-safety workflows. The official model card still acknowledges weaknesses in complex motion, long-range consistency after editing, and precise text rendering. Current quality claims are based primarily on Google's internal human-preference evaluations, with no public, reproducible benchmarks yet available for extension drift or audio-video synchronization.