Shipping Cara 3, Anam's real-time avatar model

We've shipped Cara 3, Anam's third-generation real-time avatar model. It takes audio and generates video of a face speaking it.

I'm Anam's Founding Product Engineer, working on the product experience across Lab and the SDK. The model is the research team's work; my focus is on the products people use to try it and build with it.

How Cara 3 generates a face#

In Anam's technical account of Cara 3, Ben Carr describes two stages. First, an audio-to-motion diffusion transformer converts audio into a representation of facial movement, including lip shape, gaze and head position. A separate renderer then applies that movement to a reference image to produce video frames.

That separation lets the model animate different faces without retraining for each identity. The company article covers the research and evaluation in more detail.

Where it fits into a conversation#

Face generation is one part of the response a person waits for. A conversational app also has to handle the user's input, produce an answer, generate speech and deliver the result to the browser. The time a model takes to generate its first frame doesn't measure that whole exchange.

Streaming lets stages start working on partial input, so they don't all have to wait for complete sentences or finished audio. It still leaves dependencies: the face needs audio to animate, and that audio has to come from somewhere. I cover that distinction in what 100ms actually means.

For someone building with an avatar, a useful test is to follow a complete turn: stop speaking, wait for the response, then interrupt it. Check what the browser shows and plays at each point. A face-generation timing alone can't tell you whether that interaction works well.

Try Anam#

You can try the current product in Anam Lab. For the later model release and the work around it, I've written about cara-4 and the product experience in Lab.