Generating a few seconds of striking AI video has become relatively easy. The harder problem arrives when a creator wants to continue the same scene: in the next shot a character’s face can change, the setting can shift, or the camera movement no longer matches what came before. Google’s Gemini Omni 1.1 Flash, released in late August 2026, is aimed squarely at that problem.
Ten seconds of memory
The key change is that when extending a video, the model can take up to 10 seconds of the preceding footage into account as context. Rather than guessing the next scene from the final frame alone, the model retains where the character was, what the environment looked like and which way motion was heading. Building on that continuity, a clip can be extended in 10-second segments to a total of 40 seconds.
To give creators more control, the start and end frames of a shot can now be set manually. The user decides where the shot begins and where it should finish, and the model generates the motion in between. Google says this makes camera moves, zooms, transitions and seamless loops easier to produce. A reference video of up to three seconds can also be supplied to help maintain the visual style of a character or movement.
Cheaper iteration
A practical benefit of the workflow is lower-cost experimentation. Instead of rendering every idea at high resolution immediately, creators can test a shot, prompt and composition with a 360p preview. According to Google, that preview is up to 60 percent faster than 720p output and costs roughly a third as much. Once the result looks right, the video can be rendered at 1080p and upscaled to 4K.
Taken together, these features address different parts of a single challenge: moving AI video from a collection of attractive but disconnected clips toward a continuous scene. If characters, settings and motion hold steady across several shots, and creators gain more control over the beginning, end and reference material, AI video can move beyond short clips and closer to genuine storytelling.
