bithuman 2.2.2

bitHuman avatar runtime + bundled CLI — `bithuman serve model.imx` for talk-to-your-avatar quickstart.

bithuman

This is the Python flavor of Layer 3: a platform-specific library for app developers. It wraps the Layer 1 libessence engine. For the CLI tool see docs/CLI.md.

┌─────────────────────────────────────────────────────────────┐
│ Layer 3: Platform-specific libraries (app developers)       │
│   - Python wheel       pip install bithuman    ◄──── you are here
│   - Swift package      SwiftPM Bithuman                     │
│   - Kotlin AAR         ai.bithuman:sdk                      │
│   - (future) Rust crate, JS/TS, Go, ...                     │
└─────────────────────────────────────────────────────────────┘
                          ▼ embeds
┌─────────────────────────────────────────────────────────────┐
│ Layer 2: bithuman CLI (end-user tool)                       │
│   - one cross-platform binary on macOS / Linux / Windows    │
│   - brew install bithuman · curl-pipe installer             │
└─────────────────────────────────────────────────────────────┘
                          ▼ links
┌─────────────────────────────────────────────────────────────┐
│ Layer 1: libessence engine (cross-platform C++ core)        │
│   - portable C ABI, same source on every target             │
│   - macOS · iOS · Android · Linux · Windows                 │
│   - never imported directly by app developers               │
└─────────────────────────────────────────────────────────────┘

Python bindings for the bitHuman SDK — the portable C++ avatar engine (libessence) that powers our cross-platform lipsync pipeline. The wheel ships a native pybind11 module that talks directly to libessence, so you get the same per-frame cost as our Swift and Kotlin clients with none of the GIL noise.

On an Apple M5 with 24 GB unified memory we measure ~640 FPS sustained compose (1.56 ms/frame mean, 2.03 ms p99) for a 1248×704 avatar, with ~206 MB peak RSS end-to-end. Cold load is ~14 ms for the fixture and ~400 ms for the first compose tick (lazy ONNX init).

This package is namespace-isolated from the v0 bithuman SDK; you can install both side-by-side.

Install

pip install bithuman

Status. The PyPI bithuman wheel is at v2.2.1 (2026-05-25) shipping the bundled Rust CLI + conversation brain. bithuman run is the fast-path live-avatar command; bithuman[local] adds a fully on-device brain (whisper.cpp + llama.cpp + Supertonic). The legacy low-level streaming API (AsyncBithuman.push_audio + async for ... in runtime.run()) is still exported for library users — see the legacy quickstart.

Compatibility

• Platforms: macOS arm64, Linux x86_64, Linux arm64 — all ship as wheels. Windows is tracked for a follow-up.
• Python: 3.10 – 3.13 (cp310, cp311, cp312, cp313). CPython only.
• ABI: the published wheel wraps libessence ABI v4. The libessence engine itself is on ABI v6 — that surface is currently exposed via the Swift / Kotlin / Rust bindings only. PyO3 wheel migration in flight.
• Auth: ships with live heartbeat against api.bithuman.ai baked into libessence. Avatar.load(api_secret=...) is the entry point; BITHUMAN_API_SECRET env var works too. Set BITHUMAN_UNMETERED=1 for dev / parity-test runs.

What you get

The package exposes three API tiers (all importable from bithuman):

Tier	Types	Use when…
Async	AsyncAvatar, AudioChunk, VideoControl, VideoFrame	Hosting a service / parity with legacy AsyncBithuman
Sync facade	Avatar, ComposedFrame, EP	Offline / batch / CLI rendering
Low-level	Fixture, Runtime, EP_CPU/EP_AUTO/EP_COREML/EP_NNAPI/EP_QNN	Direct C ABI access, custom audio pipeline

Error types: BithumanError (base), TokenError / TokenExpiredError / TokenValidationError / TokenRequestError / AccountStatusError (auth), ModelError / ModelNotFoundError / ModelLoadError / ModelSecurityError / ExpressionModelNotSupported (fixture), RuntimeNotReadyError.

Version info: bithuman.__version__ (Python package), bithuman.__core_version__ (linked libessence), bithuman.__abi_version__.

Quickstart (legacy AsyncBithuman — PyPI)

This is the shape of the current published wheel. It ports directly from the v0 bithuman SDK: feed PCM with push_audio, drain frames from the runtime.run() async generator.

import asyncio
from bithuman import AsyncBithuman

async def main():
    runtime = await AsyncBithuman(
        model_path="model.imx",
        api_secret="...",  # or BITHUMAN_API_SECRET env var
    ).start()

    await runtime.push_audio(pcm_16k_mono_int16_bytes,
                             sample_rate=16000, last_chunk=True)

    async for frame in runtime.run():
        # frame.bgr_image is (H, W, 3) uint8 in BGR order
        ...

asyncio.run(main())

PCM accepted is int16 little-endian bytes at 16 kHz mono. WAV / MP3 / FLAC / OGG decoding is the caller's responsibility (use soundfile).

Quickstart (low-level, C-level streaming surface)

The Rust PyO3 wheel will expose the ABI-v6 streaming pair (runtime.push_audio + runtime.pull_frame) on the same shape as the Swift / Kotlin bindings. Until it ships to PyPI, the snippet below uses the legacy Fixture / Runtime types in the published wheel.

Live avatar — bithuman run (browser + voice chat)

The wheel also ships the bithuman Rust CLI plus an embedded conversation brain (a livekit-agents worker). bithuman run avatar.imx stands up the whole stack — embedded livekit-server, libessence runtime, brain, browser player — and prints a URL you open to talk to the avatar.

pip install bithuman
export BITHUMAN_API_KEY=...     # avatar-runtime auth (https://www.bithuman.ai/#developer)
export OPENAI_API_KEY=sk-...    # the conversational brain (OpenAI Realtime)

bithuman run ~/.cache/bithuman/models/sample-avatar.imx
# → open the printed http://127.0.0.1:8088/<CODE> URL in a browser

Fully on-device — bithuman[local]

For zero-cloud operation, install the [local] extra and set BITHUMAN_LOCAL=1. No OpenAI key, no outbound network — the brain swaps to whisper.cpp (STT) + llama.cpp (LLM) + Supertonic (TTS), all in-process, all auto-downloaded from HuggingFace on first run.

pip install 'bithuman[local]'

export BITHUMAN_API_KEY=...
BITHUMAN_LOCAL=1 bithuman run ~/.cache/bithuman/models/sample-avatar.imx

Slot	Library (mobile-portable C++ core)	Default model	Disk	RAM
STT	pywhispercpp → whisper.cpp	tiny.en	77 MB	~150 MB
LLM	llama-cpp-python → llama.cpp	Qwen 2.5 0.5B-Instruct Q4_K_M	400 MB	~600 MB
TTS	supertonic → ONNX Runtime	Supertonic 3 (voice M1, 31 languages)	380 MB	~600 MB
VAD	livekit-plugins-silero	Silero	5 MB	~50 MB

Total ~860 MB on disk, ~1.5 GB RAM, ~717 ms warm load, ~1.4 s warm end-to-end (STT + LLM + TTS) on Apple Silicon. Cold start adds ~90 s once for first-run model downloads.

Optional knobs (env vars)

Var	Default	What
BITHUMAN_LOCAL	unset	=1 flips the brain to the local stack.
BITHUMAN_LOCAL_WHISPER	tiny.en	whisper.cpp model size (tiny.en / base.en / small / large-v3-turbo).
BITHUMAN_LOCAL_LLM	Qwen/Qwen2.5-0.5B-Instruct-GGUF	HuggingFace repo id of a GGUF LLM.
BITHUMAN_LOCAL_LLM_FILE	qwen2.5-0.5b-instruct-q4_k_m.gguf	GGUF file within the repo.
BITHUMAN_LOCAL_VOICE	M1	Supertonic voice preset (M1–M5 / F1–F5).
BITHUMAN_LOCAL_LANG	en	Supertonic language (31 supported: en, ko, ja, es, de, …).
BITHUMAN_INSTRUCTIONS	short default	Override the system prompt.

All three local backends have first-party iOS/Android C++ builds, so the same .gguf / .bin / .onnx model files are reusable when porting to mobile — see cpp/bindings/python/livekit_local_plugins/README.md.

CLI

A essence-render console script ships with the wheel:

pip install 'bithuman[cli]'

essence-render \
  --model ~/.cache/bithuman/models/sample-avatar.imx \
  --audio speech.wav \
  --output out.mp4

Pass --output - to stream raw BGR24 frames to stdout (handy for piping into a separate ffmpeg pipeline or a custom encoder). Other flags:

Flag	Default	Description
--fps	25	Output FPS for the MP4 container.
--quality	80	libx264 quality 1..100 (higher = better).
--ep	cpu	Execution provider hint (cpu/auto/coreml/…).
--threads	1	ORT intra-op thread count.
--no-audio	–	Skip audio muxing; produce a silent video.

Example end-to-end run (5 s sine sweep):

essence-render 0.1.0: model=sample-avatar.imx audio=sine_sweep_5s.wav ep=cpu threads=1
essence-render: loaded fixture in 14.9 ms — 1248x704 @ 25 fps, 183 clusters, 202 src frames
essence-render: composed 122 frames in 1.83s (14.96 ms/frame, 66.8 fps)
essence-render: wrote /tmp/sine_sweep_5s.mp4

(Throughput here is bounded by H.264 encode, not Essence inference. Use --output - if you want to measure raw compose speed.)

Low-level API

If you need finer control or want to swap in a custom audio pipeline, the C ABI is exposed directly:

import numpy as np
from bithuman import Fixture, Runtime, EP_CPU

fx = Fixture("model.imx", preferred_ep=EP_CPU, intra_op_threads=1)
rt = Runtime(fx)
pcm = np.fromfile("speech.f32", dtype=np.float32)  # 16 kHz mono float32
cluster_idx, bgr = rt.tick_compose(pcm, frame_idx_hint=-1)
# bgr.shape == (fx.frame_height, fx.frame_width, 3), dtype uint8

Pass the entire pcm buffer to each tick_compose call; the runtime maintains an internal cursor and advances one tick per call until the audio is exhausted.

Zero-alloc hot path (since 1.12.4)

For tight render loops, pre-allocate the BGR buffer once and pass it via out=. The runtime writes into it in place and returns just the cluster_idx. This drops wrapper overhead to within ~3 % of raw libessence (vs ~8 % for the alloc-per-tick path):

out = np.empty((fx.frame_height, fx.frame_width, 3), dtype=np.uint8)
for _ in range(num_ticks):
    cluster_idx = rt.tick_compose(pcm, -1, out=out)
    # `out` now holds this tick's frame; read it before the next call.

The same out= keyword works on tick_compose_to_size. See docs/ARCHITECTURE.md §9 for the cross-wrapper perf table.

Build from source

You need the prebuilt parent C++ archive at cpp/build/libessence.a (run the parent CMake build first), plus the runtime deps from Homebrew (onnxruntime, webp, ffmpeg, hdf5, jpeg-turbo).

cd cpp/bindings/python
uv pip install -e '.[cli,test]' --no-build-isolation

The CMake glue links the prebuilt static archive directly — it does NOT re-run the parent build, so iterate on bindings without paying the C++ rebuild cost.

Performance

Measured with tests/bench.py against the v1 compose path (audio → composited BGR frame) on Apple M5 24 GB, libessence 1.16.0:

Metric	Alloc per tick	out= reuse buffer
Steady-state mean	1.53 ms / frame	1.45 ms / frame
p99	1.66 ms	1.53 ms
Sustained throughput	655 FPS	692 FPS
Overhead vs raw libessence	+8.3 %	+2.6 %
Peak RSS (proc)	192 MB	182 MB

Wrapper overhead is within 5 % of raw libessence on the out= path; see docs/ARCHITECTURE.md §9 for the apples-to-apples methodology and the cross-wrapper comparison. Reproduce with:

scripts/bench-wrappers.sh

Linux wheels

Pre-built manylinux_2_28 wheels ship for x86_64 + aarch64 across cp310 through cp313 — 8 wheels in total, all auditwheel-repaired with the full dep tree bundled (ORT, FFmpeg, HDF5, libjpeg-turbo, libwebp, libcurl, OpenSSL).

To rebuild them locally:

# One-time: build the dep-baked Docker images (~10 min each).
docker build --platform linux/amd64 -t libessence/manylinux-x86_64:0.1 \
    -f scripts/Dockerfile.manylinux-x86_64 scripts/
docker build --platform linux/arm64/v8 -t libessence/manylinux-aarch64:0.1 \
    -f scripts/Dockerfile.manylinux-aarch64 scripts/

# Per wheel build (~2 min):
docker run --rm --platform linux/amd64 -v "$REPO":/src \
    -e PYTAG=cp311 -e ARCH_INSIDE=x86_64 \
    libessence/manylinux-x86_64:0.1 \
    bash /src/cpp/bindings/python/scripts/build-wheel-in-container.sh

Limitations

• Windows wheels not yet built — tracked for v0.2.
• The CLI's output framerate is fixed at 25 fps to match the model's internal rate. Pass --output - and pipe to your own encoder if you need temporal resampling.
• preferred_ep=COREML/NNAPI/QNN is accepted but currently no-ops to CPU in the v0.1 build.

License

Commercial. Contact [e-mail].

See also

• Root README.md — install matrix
• cpp/README.md — libessence engine internals + C ABI
• docs/CLI.md — bithuman CLI reference
• cpp/bindings/swift/README.md — Swift binding
• cpp/bindings/kotlin/README.md — Kotlin/Android binding
• docs/BUILD_AND_RELEASE.md — release flow