Veo 2 was trained on numerous video-description pairs, potentially including YouTube content, raising concerns about data usage without consent. DeepMind employs prompt-level filters and watermarking technology to mitigate risks like regurgitation and deepfakes. Alongside Veo 2, Google DeepMind announced upgrades to Imagen 3, its image generation model, enhancing image quality and prompt adherence. These developments reflect DeepMind's ongoing efforts to integrate AI capabilities across Google's ecosystem while addressing ethical and creative concerns.
Key takeaways:
```html
- Google DeepMind has announced Veo 2, a next-gen video-generating AI that surpasses OpenAI's Sora in resolution and duration capabilities, although current usage is limited to 720p and eight-second clips in Google's VideoFX tool.
- Veo 2 offers improved physics understanding, camera controls, and sharper footage, but still faces challenges with coherence, consistency, and realism in complex scenes.
- DeepMind is collaborating with artists and producers to refine Veo 2, while addressing concerns about training data sources and copyright issues.
- Google DeepMind also announced upgrades to Imagen 3, enhancing its image generation capabilities with better composition and detail rendering in various styles.