Building Anam Lab for cara-4

Today we're releasing cara-4, Anam's latest model for generating the face of an AI avatar during a live conversation. Alongside the model, I've been building the experience people use to discover a persona, create one of their own and start talking to it in Lab. My Cara 3 post explains how the model turns audio into a speaking face and where that step fits into a live conversation.

I'm Anam's Founding Product Engineer, working across the product experience in Lab and the SDK. For this release, my work included rebuilding the Personas home and creation flows, and developing the portrait controls and previews. The research team built the model, adding more expressive animation, native portrait video and support for a wider range of visual styles. Ben's launch article explains those advances.

Those changes open up more choices about what an avatar looks like and how it behaves. In Lab, the task was to bring those choices into a journey someone could follow, from an initial idea to a conversation.

Start with a conversation#

The new Personas home gives people somewhere to begin. You can browse the library, choose a persona and talk to it before deciding whether to add it to your own collection. The call opens around the avatar's video, with a portrait layout that also works on a phone.

I built that experience alongside two routes for creating a persona. Persona Design starts with a written description of who you want to create and what they should do. Digital Clone starts with a photo or webcam capture. Both lead through the choices that make the persona useful, including its voice, the knowledge it can draw on and how expressive it should be, then into a live call.

Once you've chosen a voice and given the persona a purpose, you can hear and see how they work together, then adjust what needs changing. Our experiments with customer and homepage demos explored a related question: what gives someone a reason to try the product in the first place?

Getting the portrait right#

Portrait video was a good example of how a model change affected that journey. Cara-4 can generate a portrait frame directly, giving us a format suited to a phone screen. But the framing also needs to make sense while someone is creating the avatar, before there's any video to show.

A square source image can produce different landscape and portrait crops. If the preview shows the whole source, the face can appear to change position when the call starts. I worked on previews that use the same crop as the generated video, so the still image and the live call agree.

For square captures, I added a preview that switches between the two output orientations, with resolution guidance before creation. Someone can check the framing while they still have the chance to take another photo. The API also exposes the corresponding preview images so developers can use the right one in their own interfaces.

During development, we also had a portrait control that could appear for a model version that didn't support it. I updated the control to follow the version the session would actually use, and added a useful server error for an unsupported request. This kept the choice in the interface consistent with what the system could deliver.

Choosing how the avatar performs#

Cara-4 also introduces Director Notes in beta. These let a builder choose a performance style or describe one in a short prompt. Expressivity controls how strongly the direction comes through, and cues can change the performance during a response.

For example, a language tutor might have an encouraging expression while someone practises, with a calmer delivery when explaining a correction. The controls belong alongside the persona's voice and instructions because you're shaping how the same character looks and sounds.

The model doesn't yet follow every direction precisely. The useful step is being able to try a style in a conversation and adjust it, with the controls available as part of building the persona.

Make it your own#

The visual choices have expanded too. Cara-4 supports stylised images as well as photographs. In the Digital Clone flow, the Animated 3D option restyles a portrait before it becomes a conversational avatar.

There's an animated version of me here to try:

You can explore cara-4 or create a persona in Lab. Give it a purpose, choose how it should look and sound, and try a conversation of your own.