多模態模型
Gemini Omni 1.1 Flash Extends Video Using Ten Seconds of Context and Adds First-and-Last-Frame Control
Google has promoted Gemini Omni 1.1 Flash to a stable API. Scene extension can analyze the preceding ten seconds of content and generate videos segment by segment up to a total length of 40 seconds. The new version also supports first-and-last keyframes, 360p drafts, and 4K output, although the publicly available quality data still relies primarily on Google’s own evaluations.
Google has released the stable version of `gemini-omni-1.1-flash`, bringing video generation and conversational editing together in a single Interactions API. The most substantial change is scene extension: whereas the previous version mainly referenced the final second of the source video, the new model can analyze up to ten seconds of existing content and then extend it in ten-second increments, up to a cumulative video length of 40 seconds. Applications can pass `previous_interaction_id` to continue from an earlier generation state instead of redescribing the entire scene in every turn. This is especially important for maintaining character appearance, camera movement, and narrative continuity.
The model also accepts specified first and last frames, generating a continuous shot between them. This can be used for orbiting shots, dolly or zoom movements, transitions, and looping videos. Each API call outputs 3 to 10 seconds of video at 24 FPS, with resolutions including 360p, 720p, 1080p, and 4K. According to Google, system throughput for 360p drafts can be up to 60% faster than at 720p, at roughly one-third of the cost. This makes the format suitable for generating multiple storyboard variants before upscaling the selected result. Developers can also provide a reference video of up to three seconds to borrow motion or preserve character consistency. The legacy `gemini-omni-flash-preview` endpoint is scheduled to be retired on September 30, so existing integrations must migrate to the stable model identifier.
The technical value of this update lies in shifting generative video from one-off prompts toward workflows that preserve state and support explicit boundary conditions. Products can separate drafting, review, extension, and upscaling into different cost tiers. However, “production-ready” does not mean that long-range consistency has been solved: 40-second videos are still assembled through multiple generations, so errors may accumulate from one segment to the next. The 4K option is also upscaled output and does not demonstrate native detail or physical accuracy. Google’s comparisons rely mainly on internal prompt sets and human-preference ratings; systematic third-party evaluations of seams, character drift, speech synchronization, and reproducibility across multiple turns are still lacking.