LemonSlice, a company specializing in interactive avatar video models, is launching Lemon Slice 2, a new model capable of generating real-time video. The API allows users to upload a photo and create a FaceTime-style call with a character. The team aims to move beyond chatbots and voice AI towards video avatars as the future of conversational AI.

Generating real-time video has always been a challenge, especially in overcoming the uncanny valley effect. LemonSlice's approach involves a 20B-parameter diffusion transformer that runs at 20fps on a single GPU. Techniques like causal modeling, sliding window attention, and GAN-based distillation were crucial in achieving real-time performance. They're also offering a LiveKit integration.

This technology has implications for various applications, including training simulations and role-playing scenarios. The company is actively working on improved text control for avatar movements and facial expressions. LemonSlice's release signals the growing importance of video in AI interfaces, potentially changing how we interact with AI assistants and services.

With LemonSlice 2 and its new API, developers have a powerful tool to enhance their voice agents with engaging video avatars. The company provides a playground and a coupon code for users to experiment with the technology. It will be interesting to see how developers integrate these real-time avatars.