Deep Dive
What is Veo 2 and how does it compare to OpenAI's Sora?
Veo 2 is Google DeepMind's next-generation video-generating AI, capable of creating 2-minute-plus clips in resolutions up to 4K (4096x2160 pixels). This is 4x the resolution and over 6x the duration of OpenAI's Sora, which can produce up to 1080p, 20-second clips. However, in Google's experimental tool VideoFX, Veo 2 videos are currently capped at 720p and 8 seconds.
What are the key improvements in Veo 2 compared to its predecessor?
Veo 2 features an improved understanding of physics and camera controls, producing clearer footage with sharper textures, especially in scenes with movement. It can more realistically model motion, fluid dynamics, and properties of light like shadows and reflections. Additionally, it offers enhanced camera positioning and movement for capturing objects and people from different angles.
What are the limitations of Veo 2 in video generation?
Veo 2 struggles with coherence and consistency over long durations, particularly with complex prompts. Character consistency, intricate details, and fast, complex motions remain challenging. The model also exhibits issues like lifeless eyes in animations, physically impossible facades, and blending of pedestrians and backgrounds.
How is DeepMind addressing ethical concerns around Veo 2's training data?
DeepMind uses prompt-level filters to mitigate risks like regurgitation of training data and employs its proprietary watermarking technology, SynthID, to embed invisible markers in Veo 2-generated frames. However, the lab does not offer a mechanism for creators to remove their works from existing training sets, maintaining that training on public data is fair use.
What role do creators play in the development of Veo 2?
DeepMind collaborates with creators like Donald Glover and The Weeknd to understand their creative processes and refine its video generation models. Feedback from these collaborations informed the development of Veo 2, and DeepMind continues to work with trusted testers and creators to improve the model.
What other AI model upgrades did Google DeepMind announce alongside Veo 2?
Google DeepMind announced upgrades to Imagine 3, its commercial image generation model. The new version creates brighter, better-composed images in styles like photorealism, impressionism, and anime. It also follows prompts more faithfully and renders richer details and textures. UI updates to ImageFX include chiplets for key terms in prompts, allowing users to iterate or select auto-generated descriptors.
- Veo 2 generates longer videos (2+ minutes) at higher resolution (4K) than Sora.
- Currently available in Google's VideoFX tool with limitations on resolution and duration.
- Future release on Vertex AI and integration into Google products planned.
Shownotes Transcript