Gemini Omni 1.1 Flash Brings More Control to AI Video

Introduction

AI video generation is moving beyond isolated clips toward workflows that resemble real production. Google DeepMind’s Gemini Omni 1.1 Flash is positioned around that transition. Rather than presenting only a new image-quality milestone, the update focuses on controls that developers can expose inside editing tools, storyboarding products, and creative applications. The release adds scene continuation, keyframe-style transitions, inexpensive previews, high-resolution output, and short video references.

What is new

• Scene extension: Omni 1.1 can inspect up to 10 seconds of preceding video context and continue from the point where the clip ends. Extensions are generated in 10-second increments, with a cumulative maximum of 40 seconds. A longer context window should give the model more information about the characters, setting, and narrative direction than a system that only examines the final second. The announcement, however, does not provide a broader consistency benchmark across different genres.
• First- and last-frame control: Developers can specify the beginning and ending frames of a shot and ask the model to generate the video between them. This is especially relevant to camera orbits, zooms, whip pans, transitions, and seamless loops. In practice, the interface resembles a generative version of keyframe planning rather than relying on one open-ended text prompt.
• Faster drafting at 360p: A 360p response can be used for storyboards, composition checks, and prompt iteration before rendering a final version. Google says 360p generation can be up to 60% faster and costs about one-third as much as the model’s standard 720p output. The figures are presented as system-throughput and pricing comparisons in the announcement, so real-world results may vary by workload.
• Higher-resolution delivery: Finished projects can be generated at 1080p or 4K, giving developers a path from inexpensive experiments to more polished output. The release does not specify generation times or provide detailed quality comparisons between the resolutions.
• Video references: Multimodal input can include up to three seconds of reference video. The reference can provide visual context, motion cues, or character-consistency guidance. Google’s example uses several dancer references, although the final result will depend on the source material and prompt.

Why it matters

Together, these features make the model easier to integrate into a repeatable production pipeline. Scene continuation can reduce the need to regenerate an entire sequence when only its ending needs to change. First-and-last-frame generation gives creators a clearer way to plan motion and transitions. Low-resolution previews separate creative exploration from expensive final rendering, which could be useful for editors, advertising tools, storyboard applications, and media software.

There are still important questions. The announcement focuses on capabilities and demonstrations rather than independent evaluation, failure rates, or consistency results across long and complex scenes. Developers will need to test identity preservation, motion continuity, prompt sensitivity, and the handoff between generated footage and conventional post-production. Gemini Omni 1.1 Flash is available for building through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

Google DeepMind