Essence 2 & Expression 2

bitHuman's second-generation avatar models — essence-2 and expression-2 — real-time avatar video with on-device, browser, and cloud serving, per-minute pricing, and which to choose. Available now.

Available now. bitHuman’s second-generation avatar models, essence-2 and expression-2, are live — create them today via the API (POST /v1/agent/generate) or the dashboard. The first-generation models (essence-1 and expression-1) remain fully supported, and nothing changes for existing agents or integrations. Browse the gallery to see the second generation live and talk to a demo agent.

The second-generation lineup

The two second-generation models roll out across every surface — the REST API, the embed widget, the dashboard, and the SDKs:

expression-2 — the second-generation expression engine, for stylized and universal characters (cartoons, animals, creatures, robots — and people). Audio-driven, real-time avatar video generated from a single photo: at agent creation it trains a small per-identity model (roughly 45 minutes on a training GPU), then synthesizes the entire 416×720 scene live at 20 fps — fully generated motion, not patched onto a pre-rendered base. Runs on GPU, CPU, and Apple Neural Engine.
essence-2 — the standard Essence model for photorealistic people, and the default. It animates your identity’s footage (a full-HD 1080p identity video, generated internally from your image) at ~25 fps with a distilled engine that runs everywhere — from fully on-device (iPhone, iPad, Mac, CPU — audio and video never leave your hardware) through cloud GPUs to in-browser WebGPU/WASM. bitHuman automatically serves the right way for your hardware and quality needs — you just pick essence-2.
essence-2-max — the premium Essence model: the highest-fidelity Essence renderer served directly on dedicated cloud GPUs for close-up and hero-quality output. It has no separate creation — every essence-2 creation includes it.

At a glance

	essence-2	essence-2-max	expression-2
Guide	Essence 2	Essence 2 Max	Expression 2
Family	Essence	Essence	Expression
What it is	The standard Essence 2 model — efficient distilled renderer, serves everywhere; the default	The premium model — the highest-fidelity renderer, served on dedicated cloud GPUs	Generative motion from one photo
Best for	Photorealistic humans	Photorealistic humans, close-up/hero quality	Characters: cartoons, animals, creatures, robots
Identity source	Identity video generated internally from your image	The same internally generated identity video	Single photo
Output	Identity footage animated at its native resolution (1080p driver default), ~25 fps	Identity footage, reference fidelity, ~25 fps	Fully generated 416×720 scene, 20 fps
Serving tiers	gpu · ane · cpu (auto-routed chain) · browser (WebGPU/WASM, in rollout)	gpu	gpu · ane · cpu (auto-routed chain)
On-device	Yes (CPU / Apple Neural Engine)	—	Yes (Apple Neural Engine)
Creation	Train-on-create, 500 credits (typically about 45 minutes)	Included with the combined `essence-2` creation (instant identity prep)	Train-on-create, 2000 credits (typically about 45 minutes)
Cloud	4 credits/min	8 credits/min	4 credits/min
Self-hosted	2 credits/min	4 credits/min	2 credits/min

All three keep the platform contract unchanged: push 16-bit PCM audio in, drain real-time lip-synced video frames out. The same agent code works across every surface.

Which should I choose?

Highest visual fidelity, close-up or hero content

essence-2-max. The premium Essence renderer — the highest-fidelity model, served on dedicated cloud GPUs — pick it when image quality is the whole point and 8 credits/min is acceptable. Its identity derives from the agent’s stored identity video — generated internally by every essence-2 creation.

Cost-effective at scale, or on-device

essence-2. The standard model and the default — half the cloud price of Max, and it runs on CPU, the Apple Neural Engine, and in-browser WebGPU/WASM as well as GPU — so the same agent serves from bitHuman’s cloud, your own servers, the browser, or entirely on-device. The right default for photorealistic humans, kiosks, high-concurrency deployments, and privacy-sensitive environments.

Fully generated motion from a single photo

expression-2. The most lifelike motion in the lineup — expressions and head movement are generated live from the audio rather than replayed. Creation trains a per-identity model from your photo (roughly 45 minutes); serving spans GPU, CPU, and Apple Neural Engine tiers.

Still deciding between the families (Essence vs Expression)? Start with Essence vs Expression.

How creation works

All three models are train-on-create: you create an agent once with POST /v1/agent/generate and the platform prepares that identity’s model as part of generation. Creation is asynchronous and one-time per agent — 500 credits for essence-2 (the combined creation, Essence 2 Max included) and 2000 credits for expression-2. Poll GET /v1/agent/status/{agent_id} until the status is terminal (success / ready).

How long creation takes depends on the model — the v2 models do real per-identity work, so don’t apply a short client timeout:

Model	Identity step	Typical creation time
`essence-2`	Distills a compact identity bundle on a cloud GPU	Typically about 45 minutes (up to a few hours)
`essence-2-max`	Instant prep from the internally generated identity video (seconds)	Included with the combined `essence-2` creation
`expression-2`	Trains a per-identity model on a dedicated training GPU	Typically about 45 minutes (30–60)

Creation input is a portrait image for all three — essence-2 generates a 10-second identity video from it internally (25 fps, authored to loop seamlessly, its first and last frames match), essence-2-max derives from that same video, and expression-2 trains straight from the photo. The identity image itself is produced by Seedream 5 pro (text-to-image from your prompt); an uploaded image is treated as a reference and always regenerated through Seedream 5 edit to standardize it — never used raw. The internal identity/driver video is generated by Seedance 1.5 pro. Video input is not part of the creation contract for any model and is being removed platform-wide: do not send video — as the rollout completes, a request carrying it is rejected with 400 VIDEO_INPUT_NOT_SUPPORTED before anything is billed. Details and failure modes are in each model’s guide and the Agents API.

Three optional fields shape the generated identity (they apply to every model):

aspect_ratio — 16:9 (default), 9:16, or 1:1. Drives both the generated identity image and the driver video; images are generated at 1080p.
transparency — false by default. Set true to generate the identity image on a solid green-screen background for chroma-key / transparent embedding — the character itself never uses green.
framing — portrait (default) frames head-and-shoulders; full_body shows the whole figure including the feet, for kiosk / standing-avatar layouts.

Two notes round out the creation surface:

essence-2 is a combined creation. The one 500-credit charge trains the standard Essence 2 identity bundle and makes Essence 2 Max available from the same internally generated identity video — launch with ?model=essence-2-max (or the embed-token model field) when you want the premium model. See the combined creation.
auto — classify and route. An LLM looks at your input (the image if provided, else the prompt): a photorealistic person routes to essence-2 (combined), a cartoon / animal / creature / robot routes to expression-2. It’s the default in the dashboard’s create flow; API callers must send it explicitly — an omitted model now defaults to expression-1 (Expression 1, version v1, 250 credits), flipped from the old essence-1 default on 2026-07-12. The flip is price-neutral (still 250 credits, still v1, still ungated) and never silently selects a v2 engine or a higher price. See auto.

Which family for which character? Essence 2 works best for photorealistic human figures; Expression 2 works best for cartoonish, stylized, animal, creature, or robot characters. auto applies this rule for you, and every classifiable creation records the recommended model.

The Essence 2 family requires a photorealistic human subject — an explicit essence-2* creation with a stylized or non-human input is rejected with 422 MODEL_SUBJECT_MISMATCH before anything is billed (auto routes instead of rejecting).

Already have an agent? You don’t need to re-create it — POST /v1/agent/{code}/models adds a model to it at the same per-model rates (adding expression-1 is free and instant), and GET /v1/agent/{code}/model/download downloads a generated artifact.

Create an agent with `expression-2`

import requests

import os

resp = requests.post(
    "https://api.bithuman.ai/v1/agent/generate",
    headers={
        "Content-Type": "application/json",
        "api-secret": os.environ["BITHUMAN_API_SECRET"],
    },
    json={
        "prompt": "You are a friendly product specialist.",
        "image": "https://example.com/face.jpg",
        "model": "expression",   # family: expression or essence
        "version": "v2",         # v1 or v2 → expression + v2 = Expression 2
        "aspect_ratio": "9:16",
        "framing": "portrait",
    },
)
print(resp.json())

{
  "success": true,
  "message": "Agent generation started",
  "agent_id": "A66GYD8664",
  "status": "processing"
}

The same request shape works for the other models. Select an engine with model + version — model is expression (default) or essence, and version is v1 (default) or v2, so essence + v2 is Essence 2 while expression + v1 is Expression 1. You can still pass a full engine name directly (essence-2, essence-2-max, expression-2, essence-1, expression-1) — those pass through unchanged and version is ignored, so existing integrations keep working. Omitting model entirely defaults to expression-1 (250 credits). Invalid or retired model names return 400 VALIDATION_ERROR with no credits charged; pre-rename aliases are handled as described in Naming & migration.

Then poll until the agent is ready:

curl https://api.bithuman.ai/v1/agent/status/A66GYD8664 \
  -H "api-secret: $BITHUMAN_API_SECRET"

Launch a live session

Serving works exactly like every other bitHuman agent — the model tier is resolved for you at session launch. The fastest path is the embed widget: mint a short-lived token on your backend, then drop the iframe on any page.

// SERVER — mint a token (api-secret stays server-side)
const res = await fetch("https://api.bithuman.ai/v1/embed-tokens/request", {
  method: "POST",
  headers: {
    "api-secret": process.env.BITHUMAN_API_SECRET,
    "content-type": "application/json",
  },
  body: JSON.stringify({
    agent_id: "A66GYD8664",
    fingerprint: visitorFingerprint,   // stable per-device hex — required
  }),
});
const { data: { token } } = await res.json();

<!-- BROWSER — the agent renders on its model's serving tier -->
<iframe
  src="https://bithuman.ai/embed/A66GYD8664?token=YOUR_TOKEN"
  allow="microphone *; camera *; autoplay *"
  style="width: 400px; height: 700px; border: 0;"
></iframe>

The other delivery surfaces work unchanged too — the LiveKit plugin, the REST API (/speak, /add-context), and the SDKs.

Advanced: pin a serving tier

By default the platform routes each session down the model’s serving chain (GPU → Apple Neural Engine → CPU) and overflows on capacity. For benchmarking or placement testing you can force a specific tier by appending ?model= with a force-tier slug to the session (viewer / embed) URL. A forced tier is pinned — it never overflows, and it fails loudly if that tier is unavailable:

https://bithuman.ai/embed/A66GYD8664?model=expression-2-ane

Model	Force-tier slugs
`essence-2`	`essence-2-gpu` · `essence-2-ane` · `essence-2-cpu`
`expression-2`	`expression-2-gpu` · `expression-2-cpu` · `expression-2-ane`
`essence-2-max`	`essence-2-max` (single GPU tier)

Saved links carrying pre-rename or retired slugs keep working — see Naming & migration.

Note Tier forcing is an advanced, operational surface — slugs may be adjusted as capacity evolves. For production, omit ?model= and let the platform choose.

Where each model runs

The device/runtime matrix for the second generation:

Runtime	`essence-2`	`essence-2-max`	`expression-2`
bitHuman cloud — GPU	✅ chain tier	✅ (the only tier)	✅ chain tier
bitHuman cloud — Apple Neural Engine	✅ chain tier	—	✅ chain tier
bitHuman cloud — CPU	✅ chain tier	—	✅ chain tier
Self-hosted (your servers, CPU)	✅ SDK	—	✅ (AVX-512-class CPUs)
On-device macOS / iOS (Apple Silicon)	✅ Swift SDK	— (cloud-only)	✅ Swift SDK
Browser-local (WASM/WebGPU, no server render)	Rolling out — `?render=local` renders Essence 2 in-browser (WebGPU on Apple Silicon/desktop-class GPUs, WASM fallback) as per-identity web bundles publish; the browser rendering modes ship with `essence-1` today	—	Planned (WebGPU-only)

Cloud sessions are routed automatically; on-device and self-hosted serving use the downloaded model artifact (GET /v1/agent/{code}/model/download).

A few things to know today: essence-2’s in-browser WebGPU tier renders on the viewer’s own device (audio and frames never leave the browser), with the WASM fallback covering living idle + voice while full in-browser lip-sync lands for WASM and for expression-2. Downloadable local playback via the bithuman CLI (macOS + Linux) and locally-hosted LiveKit for the v2 models are in active development — until they ship, the packaged download is for archival / self-hosted-server integration and playback runs through the cloud or the browser. And on CPU, essence-2 sustains real-time (~25 fps) while expression-2’s CPU tier is sized for offline talking-video generation rather than real-time streaming (which uses the GPU and Apple Neural Engine tiers).

Pricing

Per active minute of avatar runtime, from the pricing schedule:

Model	Cloud	Self-hosted
`essence-2`	4 credits/min	2 credits/min
`essence-2-max`	8 credits/min	4 credits/min
`expression-2`	4 credits/min	2 credits/min

Creation is one-time and per agent: 500 credits for the combined essence-2 (Essence 2 Max included — no separate Max creation), 2000 credits for expression-2, auto charges the routed model’s rate (500 or 2000), and 250 credits for the v1 models (essence-1, expression-1). Adding a model to an existing agent charges the same per-model rates. Idle, paused, or disconnected time isn’t billed. Machine-readable schedule: GET /v1/pricing.

Naming & migration

This is the one place the historical model names are documented — every other page uses the canonical names (essence-2, essence-2-max, expression-2).

essence-2-quality → essence-2-max (renamed 2026-07-10). The API accepts essence-2-quality as a deprecated alias during the migration, and server responses (supported_models, 409 messages, model downloads, talking-video job responses) may keep reporting the pre-rename family name until the rename rollout completes. New integrations should send essence-2-max.
essence-2-light was consolidated into essence-2 (2026-07-05) and is retired: you create and serve with model="essence-2", the standard model. Requests naming essence-2-light get a 400 with a hint pointing at essence-2.
Saved links and tier pins keep working. The pre-rename essence-2-quality slug pins the same tier as essence-2-max (the serving worker accepts both); the old essence-2-light-gpu / essence-2-light-cpu slugs still pin their tiers; and links carrying essence-2-light or essence-2-light-ane route to the essence-2 default chain.

Idle behavior

All three models keep the avatar naturally alive during silences, and as of 2026-07-02 their idle loops play forward-only — footage wraps from its last frame back to its first and never plays in reverse. Expression 2 additionally bakes a real-footage idle clip into every creation, so idle is the identity’s own footage rather than generated frames. Details per model: Expression 2 · Essence 2, and the expectations overview in Session behavior & troubleshooting.

Next steps

Expression 2 · Essence 2 · Essence 2 Max — the official per-model guides.
Essence vs Expression — the two model families.
Agents API — the full create → poll → serve lifecycle.
Embed widget — ship a live session in minutes.
Session behavior & troubleshooting — latency, idle, common errors.
Pricing & credits — plans, top-ups, and the full schedule.
Changelog — the GA announcement.