Concepts

The mental model behind bitHuman — one engine, one .imx format, one push/drain loop, one model family.

Start here

The second-generation family — Essence 2, Essence 2 Max, and Expression 2 — compared, with a chooser.

The single portable file that packages an avatar — where it comes from and how to inspect it.

Model guides

The standard photorealistic model — runs everywhere (GPU, Apple Neural Engine, CPU, browser), including fully on-device. The default.

The highest-fidelity tier — served on bitHuman cloud GPUs and included with every Essence 2 creation.

Stylized characters and creatures — fully generated motion from one portrait image.

Render a finished MP4 from text or audio with any model — the asynchronous output mode.

Essence 1 and Expression 1 — still supported for existing agents and the v1 API.

The push/drain pattern every SDK shares — PCM in, lip-synced 25 FPS frames out.

From a face and voice to a live talking avatar — generate, poll, resolve, drive.

The libessence engine, the SDKs that wrap it, and the ABI + hardware matrices.