![Veo]()
Image: Credit
Google DeepMind has launched Veo 2, its advanced video-generating AI model, aiming to surpass competitors like OpenAI's Sora. Veo 2 can produce videos exceeding two minutes in duration at resolutions of up to 4K (4096 x 2160 pixels)—a notable improvement over Sora’s capabilities of 20 seconds at 1080p.
![Veo 2]()
Image: Credit
Veo 2 is only available in Google's VideoFX, an experimental tool currently capped at 720p and eight seconds. DeepMind plans to gradually expand access, with a broader rollout to developers via Vertex AI in the future.
What’s New in Veo 2?
Veo 2 has significant improvements in generating sharper textures, realistic motions, and more accurate lighting, including cinematic effects and complex dynamics like liquids and reflections. The model also features enhanced camera control, allowing precise movements and different perspectives in generated videos.
Despite advancements, Veo 2 still struggles with “coherence and consistency,” particularly with complex motions and maintaining character details over extended prompts. DeepMind is actively collaborating with artists and creators to refine the model further.
Safety and Training
Veo 2 was trained on extensive video-description data, though DeepMind did not disclose exact sources, with speculation pointing to platforms like YouTube. DeepMind uses its SynthID watermarking technology to embed invisible identifiers into generated frames to mitigate deepfake risks.
Imagen 3 Upgrades
Alongside Veo 2, DeepMind announced updates to Imagen 3, its image-generation model. Available via Google’s ImageFX, Imagen 3 now produces brighter, more detailed visuals across styles like photorealism and anime. New interface features allow users to enhance prompts with suggestions for improved outputs.
As Google DeepMind continues to develop its AI tools, these advancements position the company as a leader in the competitive AI video and image-generation space.