Genie 4 opens in Project Genie: Google's world model now takes live prompt edits and makes its own sound
On Nov 18 Google opened Genie 4 to eligible Google AI Ultra subscribers inside Project Genie. You can type a new prompt while you walk through a generated world and the world changes, and the audio (ambience, music, dialogue) is generated alongside the picture. It's a research preview, and the output is 720p, but it's the clearest sign yet that world models are heading toward previs and blocking.
What shipped. Google's post (11:15 AM MST Nov 18) puts Genie 4, its newest world model from Google DeepMind, inside Project Genie, the browser prototype for making and walking through interactive worlds. Until now Genie 4 was limited to trusted testers (since late October, per The Verge). The headline feature is editing mid-session: while you move around, type a prompt to change the weather, drop in a character or redirect what's happening. Per Google, it's the first Genie that pairs the visuals with real-time native audio, from ambient sound to character dialogue.
What it outputs. Per the model card, Genie 4 takes text, images and live keyboard/mouse input and streams 720p video at up to 24 fps with synced audio. It's built on Gemini Omni Flash, with Gemini handling reasoning and steering so the world stays coherent. Everything Project Genie generates carries a SynthID watermark.
Around it. A new Project Genie agent builds a world from a conversation, including a look-setting first pass and a custom score. Gemini can also ride along in a session, suggesting where to go next or triggering events. Genie 3 stays in the model picker and got its own upgrades: sessions up to 3 minutes (up from 60 seconds), chaining worlds together, seeding a world from a 360-degree panorama, and live in-world events. A community gallery lets people explore and remix each other's worlds.
The limits Google lists. Genie 4 holds about a minute of continuous interaction. Details and object permanence drift over longer sessions or after several prompt edits. Complex physics, legible text (unless it's in your description), real-world locations and audio accuracy are all called out as unsolved. Treat it as a sketchpad, not a render engine.
Why video people should care. DeepMind itself names filmmaking as a use. Realistically that means previs: walking a virtual location, testing a camera move or weather change, and hearing a rough soundscape before you spend credits on a video model. Screen-capture it as reference footage, or pull stills into an image-to-video workflow. Access is the catch for now: eligible AI Ultra subscribers only, with a waitlist for AI Pro.
Signals: official Google post (about 2,200 likes within a day) plus a tracked-curator post from @RoundtableSpace on Nov 19, so it's HOT by our rule. Coverage from The Verge and The Decoder.
Related tools on ReelStack
More from the Lab
- OpenMotion 0.2.0 adds seven studios for images, voice, video and 3D; plus LTX's SDR-to-HDR comparison, Kling Image 4.0 in ComfyUI and HyperFrames 0.9.72
- Comfy's lip-sync dubbing workflow redraws only the mouth: one shot into any language with ElevenLabs and MiniMax H3; plus Genie 4 and HeyGen Decision 1 in ComfyUI
- HyperFrames team says reverse-any-video beats REA-Video, Comfy Cloud adds custom Builds, and a test of camera-lens prompts
- ComfyUI v0.45.0 for video makers: RelayVSR video upscaling, H3-Image, Seedream 5.5 and Breeze TTS 2 nodes
Topics: Audio, voice & music