Google's new AI creates stunning 4K videos using your prompts
What's the story
Google's DeepMind divison has unveiled Veo 2, a new video-generating AI model that aims to beat OpenAI's Sora in terms of video resolution and duration.
The advanced model can generate clips longer than two minutes at up to 4K resolution, a major upgrade over Sora's capabilities(up to 1080p, 20-second-long clips).
Currently, Veo 2 is accessible through Google's experimental tool, VideoFX. Although it is behind a waitlist, Google has announced plans to broaden access to more users later this week.
Future plans
Veo 2's integration into Google's ecosystem
Eli Collins, VP of Product at DeepMind, has revealed plans to make Veo 2 available on Google's Vertex AI platform as it becomes ready for large-scale use.
User feedback will guide further development and integration into Google's ecosystem.
The advanced model can generate videos from text prompts or text with a reference image and offers improved "understanding of physics" and camera controls for clearer footage.
Advanced features
Veo 2's realistic modeling and cinematic capabilities
DeepMind claims that Veo 2 can realistically model motion and fluid dynamics while handling light properties such as shadows and reflections.
The model also supports different lenses and cinematic effects with nuanced human expressions.
Despite these advancements, challenges remain in coherence, consistency, and realism over complex prompts and long durations.
Model refinement
DeepMind's collaboration with artists and data extraction policies
DeepMind worked with artists such as Donald Glover and The Weeknd to improve its models.
The company has trained Veo 2 on thousands of video-description pairings but hasn't revealed the exact data sources.
While DeepMind offers tools for webmasters to prevent data extraction, it doesn't let creators remove data that has already been used for training.
Safety protocols
DeepMind's safety measures and Imagen 3 upgrades
DeepMind employs prompt-level filters for content safety and SynthID watermarking tech to reduce deepfake risks.
Google's indemnity policy won't cover Veo 2 until it's generally available.
Along with Veo 2, DeepMind also announced upgrades to Imagen 3, its image generation model. The new version provides enhanced image composition and detail across different styles such as photorealism and anime.