Kandinsky WM 1.0: Generative Models for Physical AI

Sber has open-sourced Kandinsky WM 1.0, a set of three specialized video generation models for Physical AI: K5-AV for autonomous vehicles, K5-RO for robotics, and K5-PH for scenes with complex physics. The models are based on the 2-billion-parameter Kandinsky 5.0 Video Lite and operate in image-to-video mode, taking a first frame and a text prompt to generate a continuation. Each model was domain-adapted on specific datasets: 1.5M videos from NVIDIA PhysicalAI Autonomous Vehicles, 2.2M robotics videos from various sources including GenRobot 10Kh RealOmni and AgiBot World Beta, and 1.6M physics videos with 1.3M from pre-training and 300K manually curated. After domain adaptation, the models underwent a second training stage: RL with a reward model for K5-AV and K5-RO, and SOAR for K5-PH. The release includes code and weights under the MIT license. The article also highlights the long-tail data problem in Physical AI, where rare scenarios are critical but difficult to collect, and notes that current video generation models often lack physical fidelity, necessitating domain-specific training and post-training.