Gemini Omni 1.2 Flash

What Gemini Omni 1.2 Flash is good at, and where it isn't

Gemini Omni 1.2 Flash keeps the idea that made Omni 1.1 useful, one model that creates, edits and extends video and generates the sound in the same pass, and removes the limit people hit first: length. Google released it on November 5, 2026 with clips of up to 15 seconds, closer reference following, cleaner speech and a lower price. From December 3, 2026, Google routes every Omni 1.1 request to 1.2, so this is the version to build on.

The first flagship change is 15-second generation. Duration goes from 3 to 15 seconds in one-second steps, which is enough for a complete social spot: an opening, the product in use and a closing hero shot, scored and in sync, without stitching three clips and hoping the lighting matches. In our test a 15-second vertical clip came back in one take with continuous audio.

The second is references that hold. Add up to 10 images (a character from several angles, a product, a location) and up to three 3-second motion clips. Google built 1.2 to keep faces, garments, packaging typography and brand colors consistent from the first frame to the last, with less morphing during camera moves. That makes it practical for virtual try-on from front and back product shots, catalog spins, and ads built from existing character sheets.

The third is editing as a loop. Upload a clip of up to 15 seconds and describe the change: a striped tent becomes an oak forest, a sunny street turns to rain, a hoodie becomes a leather jacket. The actors, framing and timing stay. Because the output is just another clip, you can edit it again, which is how Google demonstrates its "multi-turn" editing, and how you produce several placements from one take. Extend works from the other end: it follows the last 10 seconds of any clip and adds 3 to 10 seconds with matching sound.

Camera direction improves too. Long continuous moves (an FPV flight through a doorway, an orbit, a slow zoom out) come back with noticeably fewer unwanted jump cuts than in 1.1.

Where it's weaker: framing is 16:9 or 9:16 only, a last frame caps the clip at 10 seconds, and audio is always on; you can steer it but not disable it. 1080p and 4K are upscaled from the native 360p and 720p frames. Google's safety filters are strict and photos of some recognizable people are refused, so for putting a specific person into existing footage MiniMax H3 is the better tool. For takes longer than 15 seconds or square framing, Wan V3.0 goes to 30 seconds; for edits of clips longer than 15 seconds, use Seedance V2.5.

A reasonable mental model: make Gemini Omni 1.2 Flash the default for clips with sound that you expect to revise, because the same model can edit and extend what it made. Reach for MiniMax H3 to swap a person into footage, Wan V3.0 for long or square takes, Seedance V2.5 for multi-shot sequences, and Kling V4 when you want Kling's look and 1080p native output.