XGENlabs introduced the interactive model XGEN-JING

Team XGEN-JING released a namesake model for generating first-person video and audio. The system is designed to simulate navigation and interaction with objects in real-time.

The architecture is based on the basic MiniMax-H3 model. Developers provided the JING-Flash-v1 version, which is optimized for low-latency operation. The model supports keyboard control, allowing the user to directly influence the trajectory of movement and actions of the virtual agent in the frame.

To generate content, the system uses reference images as context. The project repository publishes the full code for inference, as well as a set of instructions for composing prompts to control the model's behavior. The entire toolkit is available for local deployment and testing.

The technology is focused on tasks that require egocentric vision, such as robotics training or creating interactive environments. Using JING-Flash-v1 allows generating sequences where the audio track is synchronized with visual changes caused by user actions.

The release expands the capabilities of generative systems in the field of interactive video, moving from passive viewing to active environment management. This simplifies the creation of training datasets for autonomous agents operating in first-person conditions.