Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

2.3 KB
Raw Blame
# LongCat-Video model dependencies, installed under constraints.txt.
#
# Mirrors upstream's requirements.txt + requirements_avatar.txt
# (github.com/meituan-longcat/LongCat-Video) with four deliberate differences:
#
#   * torch / torchvision / torchaudio install separately from the cu124 index,
#     before this file, so the CUDA build is the one pip resolves against.
#   * flash-attn installs separately from a pre-built wheel. Upstream's plain
#     "flash-attn==2.7.4.post1" is an sdist that compiles for ~40 minutes and
#     needs nvcc, which the CUDA *runtime* base image does not ship.
#   * opencv-python is swapped for opencv-python-headless: the GUI build pulls
#     an X11 stack no server needs.
#   * dropped from requirements_avatar.txt: libsndfile1 (a name squat on PyPI;
#     the real dependency is the apt package, installed in the Dockerfile),
#     tritonserverclient (unused), openai (only imported by the optional prompt
#     enhancer, which this worker does not call), nvidia-ml-py, tzdata, cffi and
#     chardet (nothing on the avatar path imports them), and streamlit/pyarrow
#     (the run_streamlit.py demo UI only).

# Core video pipeline.
numpy==1.26.4
transformers==4.41.0
diffusers==0.35.1
loguru==0.7.2
einops==0.8.0
ftfy==6.2.0
psutil==6.0.0
av==12.0.0
opencv-python-headless==4.9.0.80
imageio==2.37.0
imageio-ffmpeg==0.6.0

# Avatar path: audio loading, loudness normalisation, and the ONNX vocal
# separator that run_demo_avatar_single_audio_to_video.py imports at module
# scope (audio_separator.separator.Separator).
librosa==0.11.0
soundfile==0.13.1
soxr==0.5.0.post1
scipy==1.15.3
pyloudnorm==0.1.1
onnx==1.22.0
audio-separator==0.30.2
# The runtime that executes the separator's Kim_Vocal_2.onnx. audio-separator
# only pulls one in through its [cpu]/[gpu] extras, so a plain install leaves
# `import onnxruntime` (separator.py, module scope) unresolvable and every
# avatar job dies with ModuleNotFoundError before the first frame. The GPU build
# provides the same `onnxruntime` module. 1.20.x specifically: it is the last
# onnxruntime-gpu line built for any CUDA 12.x with cuDNN 9 (1.21 and later
# require CUDA 12.8), which is what this image's CUDA 12.4.1 runtime base and the
# cuDNN 9 that torch 2.6.0+cu124 loads provide. check_pins.py asserts the version,
# the CUDA execution provider, and that the separator imports.
onnxruntime-gpu==1.20.2