Google DeepMind has released its video-generating AI model, Bio 2. This model can produce videos longer than 2 minutes at up to 4K resolution, and is 4 times and 6 times higher in resolution and duration, respectively, than OpenAI's Sora.
Bio 2 generates videos based on text prompts or images, and has greatly improved physical movement and camera control. It can precisely manipulate the virtual camera to recognize objects and people from various angles. It has improved realistic physical reproduction capabilities, such as fluid dynamics and light properties. It can also implement cinematic techniques and subtle human expressions.
Google emphasized that the Bio 2 is superior to Sora in understanding the physical laws of the real world and human movement. In its own test results, it also scored higher than competing models in terms of 'prompt compliance' and 'overall preference'. However, some technical instabilities still exist, such as finger count generation.
Bio2 is currently available to some users through the experimental video production tool, VideoFX, and will be integrated into the developer platform, Vertex AI, next year. Google plans to improve the model through various feedback.
