Google AI Releases Gemini Omni 1.1 Flash: 40-Second Scene Extension, First/Last Frame Control, and 4K Upscaling

Google AI Releases Gemini Omni 1.1 Flash: 40-Second Scene Extension, First/Last Frame Control, and 4K Upscaling">

Google has released Gemini Omni 1.1 Flash ( gemini-omni-1.1-flash ), a production update to its native multimodal video generation and editing model. The release moves Omni from a capable generator to a directable one: scene extension now reads up to 10 seconds of prior context instead of a single final frame, first and last frames can be pinned to control camera movement, drafts render in 360p at a third of 720p cost, finals upscale to 4K, and video clips can be passed as references for character consistency.

Gemini Omni Flash is built on three properties Google distinguishes from prior video models: native multimodality (text, image, audio, and video processed together), conversational editing through the Interactions API , and world knowledge inherited from Gemini. Editing is stateful — you pass previous_interaction_id and the model applies your change while preserving what you did not mention, without re-uploading the prior video.

It is available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform , with Adobe, Figma Weave, GMI Cloud, and Runway already named as production users.

Omni 1.1 analyzes up to 10 seconds of prior context when continuing a clip. Google states that previous models referenced only the final second. Extensions run in 10-second increments to a cumulative 40 seconds , and the model generates a 3–10 second continuation per call. Some final frames of the input are edited to make the seam continuous.

The constraints are specific. Extension appends to the end of a clip only — no prepending, no mid-clip insertion. Uploaded input videos must be 10 seconds or shorter, unless you are extending a model-generated video in multi-turn. You cannot add new dialogue when extending an uploaded video where someone is speaking; spoken dialogue is supported in multi-turn extension via previous_interaction_id .

You can now supply a first and last frame and have the model generate the continuous video between them, which is the mechanism behind orbits, dolly-zooms, and seamless loops. Prompts bind media to roles with tags: , , , and .

Video references accept a maximum of three clips, up to three seconds each , and work best for likenesses. Audio inside a video reference is ignored. Reasoning across multiple videos is not supported and may degrade output.

The resolution parameter in response_format takes 360p , 720p (default), 1080p , and 4k , with the top two upscaled. Google reports 360p previews generate up to 60% faster and at a third of the cost of 720p, based on system throughput of 360p versus 720p. That makes the draft-then-upscale loop the intended production pattern: iterate cheaply, render once.