Pika Labs launches real-time video chat with an AI avatar version of yourself in meetings

The AI video and image generation startup Pika Labs has introduced real time video chat for AI agents, allowing live face to face conversations with AI avatars through its "Pika AI Self" concept (basically an AI generated version of yourself in different visual styles). Users can already try it by inviting a Pika AI Self into a Google Meet call. While the feature is currently tied to Pika’s own product, the company has also released a GitHub Skill so developers can connect other agents as well.

The update is powered by PikaStream1.0, which Pika describes as a real time visual engine built for identity consistent talking avatars in live calls. The company says it can stream video at 24 frames per second with about 1.5 seconds of latency on a single H100 GPU, which marks a clear shift from more traditional AI video systems that still work with longer delays. Pika also says the system can preserve memory and conversational context, while letting users change identity data during a session without restarting the stream.

The company also says it trained the model on a large filtered dataset of about 10 million pre training clips and 2 million supervised clips, while using a parallel audio and video pipeline so generation can begin as soon as audio input is available.