Anam lets its avatars' acting direction be written

Anam Cara-4 introduces Director Notes to dictate the behavior of a virtual interlocutor. What this direction layer changes for practical use.

Anam is pushing its real-time conversational avatar model forward with Cara-4. The core innovation sounds like something straight from a film set: Director Notes, an acting direction layer that lets you choose how the avatar performs the conversation.

The concept stems from real-world observations. A distressed patient should not behave like a confident sales coach, and an angry customer requires a different presence than a benevolent tutor. Users can therefore select a tone from presets, or write it as a free-form instruction like "maintain steady eye contact, stay calm, with restrained movements, as if supporting someone through a difficult conversation." An expressiveness slider then adjusts the intensity of the performance. Crucially, this direction is dynamic: over the course of an exchange, the language model can slip in acting cues while delivering its response, shifting the avatar from curious to worried, then to warm, all within a single turn of speech.

The publisher is unusually candid about the limitations: compliance with instructions is not yet at the desired level, and the avatar cannot perform the full range of gestures a developer might want to direct, such as typing on a keyboard while looking for information.

The rest comes down to production craftsmanship. Rendering moves to 1152 × 768, with portrait format treated as a first-class citizen rather than a cropped landscape. Custom avatars now accept non-photographic images: animated 3D characters, manga-style drawings, or illustrations. On the operational side, Anam claims a median connection time of under a second and a generation latency of about 150 milliseconds, with zero data retention and optional regional data residency. Finally, avatars can be invited with a single click into a Zoom, Microsoft Teams, or Google Meet call.