Genie 4 features, without the hype
Four capabilities separate Genie 4 from a video generator: real-time steering, physics and human realism, native audio, and complex event simulation.
Each is described here the way DeepMind describes it.
The four capabilities
Genie 4's feature set is best understood as four interlocking abilities, not a list of independent tricks. Steering needs physics to respond; audio needs the scene to know what it is.
- Real-time steering — Natural-language edits applied mid-session without a restart.
- Physics and human realism — Aerodynamics, collisions, and believable character motion.
- Native audio — Ambience, music-like sound, and reactive dialogue generated with the visuals.
- Complex event simulation — Multi-step causal behaviour for training and pre-visualisation.
Google has not published benchmark numbers for any of these. The strengths below are qualitative claims from the announcement, not measured results.
What each capability actually means
Translating the announcement into practical terms.
Real-time steering
Change weather, reposition the camera, or add an object or person. Gemini rewrites the prompt to fit the current context and keeps the scene advancing between your inputs.
Physics and human realism
Aerodynamics, collisions, and natural character movement with reactions. This is what makes objects feel like they occupy space.
Native audio
Setting-appropriate ambience plus interactive character dialogue that adapts live — generated with the scene, not layered on afterwards.
Complex event simulation
Gemini reasons through multi-step outcomes, modelling causality and procedural behaviour. This is the capability robotics and agent training depend on.
SynthID watermarking
Every generated video and audio track carries an imperceptible watermark designed to survive re-encoding.
Promptable world events
Inject a scripted moment into a live session — a storm, a door opening, a crowd arriving — without cutting to a new shot.
Steering in practice
What changes when the world is interactive rather than rendered.
The scene does not restart
In a text-to-video workflow, changing the weather means generating a new clip and accepting a new look. With Genie 4, the instruction is applied to the running scene, and everything else stays put.
Context is preserved across edits
Because Gemini rewrites your instruction against the current scene state, a short command like "make it rain" can be applied without re-describing the location, the camera, or the characters.
But the budget is finite
Google notes that detail and object permanence drift after many edits. Practically: every intervention costs some consistency, so fewer, larger edits hold up better than dozens of small ones.
Audio is generated, not added
Genie 4 produces sound as part of the world state. That changes what a prompt should contain — and why "silent" is now an explicit choice rather than a default.
Ambience
Setting-appropriate background sound — surf, traffic, wind, room tone — matched to the place.
Music-like sound
Tonal, musical material generated to fit the scene rather than licensed from a library.
Interactive dialogue
Characters can speak, and their lines adapt as the scene develops.
Prompt it explicitly
Naming sounds in the prompt measurably changes the result; leaving audio unmentioned leaves it to the model.
Complex event simulation
The capability that makes Genie 4 interesting for robotics and agent research: modelling consequences over multiple steps, not just the next frame.
Feature-level comparison
Where Genie 4 stands relative to its predecessor and to conventional video generation.
| Feature | Genie 4 | Genie 3 | Text-to-video |
|---|---|---|---|
| Real-time interaction | Yes, with language edits | Yes, navigation plus events | No |
| Generated audio | Native ambience and dialogue | Not described | Usually added in post |
| Physics modelling | Aerodynamics, collisions, character motion | Present but less emphasised | Per-frame inference |
| Multi-step events | Via a Gemini agent loop | Promptable events | Not supported |
| Watermarking | SynthID on video and audio | SynthID on video | Varies by provider |
| Public availability | Research preview, AI Ultra, API partners | Limited research preview | Widely available |
Genie 3 figures come from its August 2025 announcement; Genie 4 details come from the October 2026 announcement.
Scroll the table sideways to see all columns.
Feature questions
Clarifying what the announcement does and does not claim.
Put the features to work
Compose a prompt that exercises steering, motion, and audio — then generate it in a live tool.