What’s New: World Model, Conversational Editing, Audio-In, SynthID

Four things make Omni different from every video model before it.🪐

1. The world model. Omni understands the physical boundaries of reality — gravity, kinetic energy, fluid dynamics. The “marble test” (a marble running a chain-reaction track, every collision physically correct and individually audible) is the proof. Better physics intuition, not a perfect simulator — but a real leap over the previous generation.💬

2. Conversational, turn-by-turn editing. Instead of rewriting a full prompt for every change, you talk to the video: “make the violin invisible,” “change the camera angle.” Characters stay consistent, physics hold, and the scene remembers what came before. Reliable up to about 4 turns before drift sets in. 🎧

3. Audio as input. Feed Omni a music track or voiceover and it reasons across the audio to generate matching visuals — pacing, emotion, and beats lined up. Most models take one input type; Omni combines video + image + audio at once. That’s the “omni” in the name.🛡️

4. SynthID on every clip. Every output carries an imperceptible SynthID watermark, verifiable in the Gemini app, Google Search, and Chrome. Know this: it’s a provenance signal, not DRM — standard re-encoding (color grade, grain, upscaling) tends to disrupt it.⚠️

Held back at launch: editing what people say in a video (speech editing) was built but restricted — read as a deepfake hedge ahead of the 2026 US elections.