Krea Releases Open-Source Real-Time Video Model

Krea has introduced a new open-source model designed for real-time video generation. Named Krea Realtime Video, the model is available under a permissive Apache 2.0 license, allowing for broad commercial and research use in interactive applications.

The model's key innovation lies in its efficiency. It is a 14-billion-parameter text-to-video generator that was created using knowledge distillation. According to the company's release card on Hugging Face, it was trained on the outputs of a larger, more powerful "teacher" model called Wan2.1-T2V-14B. This technique allows the smaller model to achieve high-quality results with significantly lower computational requirements, enabling its real-time performance.

This release is significant because most state-of-the-art video models are kept behind closed APIs. By providing an open and performant alternative, Krea empowers developers to build novel video tools, creative applications, and new user experiences without being tied to a proprietary service. This could accelerate innovation in the generative video space by lowering the barrier to entry for real-time creation.

Model at a Glance

• Size: 14 billion parameters
• License: Apache 2.0
• Capabilities: Real-time text-to-video and video-to-video
• Architecture: Distilled from a larger foundation model