Veo 2

Veo 2 is Google DeepMind's second-generation text-to-video model, announced December 16, 2024 as the answer to OpenAI's Sora. It generates realistic video clips from text prompts (or text plus a reference image), with theoretical output up to 4K resolution and two minutes long — roughly four times the resolution and six times the duration Sora offered at the time — plus cinematic controls over lenses, angles, and effects.

Its headline improvements over the original Veo were a deeper understanding of real-world physics — fluid dynamics, light behavior, nuanced human expression — and better frame-to-frame consistency. Like Imagen 3, its outputs carry SynthID invisible watermarks as a deepfake mitigation.

Veo 2 is closed: it launched inside Google's experimental VideoFX tool (waitlisted, clips initially capped at 720p and 8 seconds) and later came to Vertex AI and the Gemini app. No weights are published.