Expression 1

Official guide to expression-1 — bitHuman's first-generation expressive avatar model: audio-driven facial animation from a portrait, GPU-only, how it is served and self-hosted, what it costs, and what the .imx artifact contains.

Note — first generation, and still maintained. expression-1 is bitHuman’s original expressive model and is not deprecated: it is what ?model=expression serves, it is the default model when /v1/agent/generate is called with no model, and thousands of live agents run on it today. For new expressive work the recommended model is expression-2.

What it is

expression-1 animates a face from a portrait image rather than replaying a pre-recorded identity: you give it audio, and it generates the facial motion to match. That is what makes it expressive, and it is also why it needs more hardware than essence-1.

-1 is naming symmetry with Expression 2, which is a different engine — not an earlier version of the same one.

Where it runs

GPU only. That is an explicit product decision, not a gap: expression-1 is served on bitHuman’s cloud GPUs and has no CPU, browser or on-device build, so a missing Apple, browser or Android artifact for this model is correct.

See Where each model runs for the per-model matrix. Nothing here restates it.

How to self-host it

One published route, and it is a container: sgubithuman/expression-avatar on Docker Hub, which needs an NVIDIA GPU. The self-hosted deployment guide has the run instructions, the digest to pin and the licence terms.

There is no Apple on-device route for expression-1. The Swift package we publish exposes bitHumanKit, BithumanEngineProtocol and Expression2; there is no Expression 1 engine type in it to import. If you need an avatar that runs on a Mac or an iPhone without a GPU server, use essence-1 or essence-2 instead.

What it costs

Billed per live minute at the first-generation expressive rate, with a one-time agent-creation charge. All numbers live on one page: Pricing & credits. The live rate card your key sees is GET /v1/pricing.

What the artifact is

expression-1 shares the .imx container and the same public store as essence-1 — the two first-generation models have one artifact format between them. You fetch it the same way:

curl -H "api-secret: $BITHUMAN_API_SECRET" \
  https://api.bithuman.ai/v1/agent/<CODE>/model/download

★ Not every expression-1 agent has one. The model can animate from a portrait without a per-identity file, so an agent created that way has nothing to download and the endpoint answers unavailable for it. That is the designed behaviour, not a broken agent: if you need a downloadable artifact, check the response rather than assuming one exists.

The engine weights are a separate thing and are not part of your download — they ship inside the self-hosting container image.