JoyAI-VL-Interaction

Why we included this project

JoyAI-VL-Interaction pairs an 8B vision-language model with a real-time streaming stack, so it can watch a webcam or livestream and respond in under a second rather than waiting for the whole clip to finish. That makes it a practical fit for assistants that have to react to live video, like safety monitoring, kitchen or workshop oversight, or any hands-busy scenario. The repo bundles the model weights, training recipe, time-aligned interaction data, and serving code in one place, so you can reproduce the system and adapt it to your own setup. It also handles the harder judgment calls by handing subtasks to a background model or agent while it keeps watching the stream. For developers who want a working, low-latency video-understanding system instead of a research-only demo, this is a solid place to start.