Models

TruGen builds and runs its own foundation models. Two families live here:

The Huma family — avatar renderers that turn audio into a photorealistic face. Two are in production today: Huma-2 for real-time conversations, and Huma-1 for high-expressiveness renders.
Hawkeye-1 — the Vision Action Recognition Model that gives agents real-time perception of the user, their environment, and their emotional state.

Together, Huma renders the face; Hawkeye reads the room.

The Huma family at a glance

Model	Render pipeline	Best fit
Huma-2	≤ 500 ms audio → video	Real-time conversational agents. Default renderer.
Huma-1	Higher — optimised for expression fidelity	Async video generation, high-expressiveness renders

The two production models take different design trade-offs. Here’s how they compare across the dimensions that matter for real deployments.

Dimension	Huma-2	Huma-1
Primary use case	Real-time conversation	Async video generation, expressive renders
End-to-end conversational latency	≤ 800 ms	Higher — not tuned for throughput
Render pipeline (audio → video)	≤ 500 ms	Longer per frame — optimised for expression
Rendering approach	Gaussian Splatting with audio-driven facial deformation	First-generation neural avatar renderer
Source video required for custom avatars	2–3 minutes controlled recording	Longer, more constrained capture
Model footprint	Leaner, faster inference	Larger, more expressive per-frame
Runtime	Two-phase (preload + inference)	Single-phase render
Expression fidelity	High — natural facial motion, gaze, micro-expressions	Highest — dense micro-expression and emotional nuance
Audio robustness	Robust across recording conditions	Tuned for clean audio input
Availability	Default on all new agents	Available on legacy agents

Whichever Huma model you use, it sits at the end of TruGen’s conversational pipeline:

User audio → STT → LLM → TTS → Huma → Video output
                        ↑
                    Hawkeye-1 (vision context, in parallel)

Full breakdown in the Realtime Pipeline.

Real-time avatar renderer — architecture, characteristics, and how it renders your agent.

First-generation expressive renderer.