Files
product/environments.md
T
2026-04-08 14:38:04 +08:00

6.3 KiB
Raw Blame History

已配置环境与已下载模型清单

最后更新:2026-03-25 机器:WSL2, RTX 5090, CUDA 12.9


一、Conda 环境总览

conda env list
环境名 路径 用途 磁盘占用
LivePortrait ~/miniconda3/envs/LivePortrait 头部动作驱动 8.6GB
MuseTalk ~/miniconda3/envs/MuseTalk 唇形同步(主力) 11GB
latentsync ~/miniconda3/envs/latentsync 唇形同步(备用) 9.2GB
condiff-train-hair ~/miniconda3/envs/condiff-train-hair 其他项目(未动)
onediff ~/miniconda3/envs/onediff 其他项目(未动)
vllm-qwen3 ~/miniconda3/envs/vllm-qwen3 其他项目(未动)
py310 ~/miniconda3/envs/py310 通用 Python 3.10

二、LivePortrait 环境

激活conda activate LivePortrait

关键包版本

版本
torch 2.12.0.dev20260324+cu128nightly
CUDA 12.8

已下载权重

位置:~/work/LivePortrait/pretrained_weights/(总计 1.2GB

pretrained_weights/
├── insightface/
│   └── models/buffalo_l/
│       ├── 2d106det.onnx          (4.8MB)   人脸关键点检测
│       └── det_10g.onnx           (17MB)    人脸检测
├── liveportrait/
│   ├── base_models/
│   │   ├── appearance_feature_extractor.pth  (3.3MB)
│   │   ├── motion_extractor.pth              (108MB)
│   │   ├── spade_generator.pth               (212MB)
│   │   └── warping_module.pth                (174MB)
│   ├── retargeting_models/
│   │   └── stitching_retargeting_module.pth  (2.3MB)
│   └── landmark.onnx                         (110MB)
└── liveportrait_animals/
    ├── base_models/               (同上,动物版)
    └── base_models_v1.1/
        └── appearance_feature_extractor.pth  (3.3MB)

来源hf download KlingTeam/LivePortrait --local-dir pretrained_weights


三、MuseTalk 环境

激活conda activate MuseTalk

关键包版本

版本 备注
torch 2.12.0.dev20260324+cu128 nightlysm_120 必须
mmcv 2.1.0 从源码编译,需要 nvcc
mmdet 3.3.0 ≥3.3.0 才兼容 mmcv 2.1.0
mmpose 1.3.2 ≥1.3.0 才兼容 mmcv 2.1.0
transformers 5.3.0 ≥4.45.0 才兼容 huggingface_hub 1.x
huggingface_hub 1.7.2 CLI 命令名为 hf
diffusers 0.32.2

已下载权重

位置:~/work/MuseTalk/models/(总计 5.5GB

models/
├── musetalkV15/
│   ├── unet.pth                   (3.2GB)   主模型
│   └── musetalk.json                        模型配置
├── sd-vae/
│   ├── diffusion_pytorch_model.bin (320MB)  VAE 解码器
│   └── config.json
├── dwpose/
│   └── dw-ll_ucoco_384.pth        (389MB)  姿态估计(DWPose
├── whisper/
│   ├── pytorch_model.bin          (145MB)  音频特征提取
│   ├── config.json
│   └── preprocessor_config.json
├── syncnet/
│   └── latentsync_syncnet.pt      (1.4GB)  唇形同步评估
└── face-parse-bisent/
    ├── 79999_iter.pth             (51MB)   人脸解析
    └── resnet18-5c106cde.pth      (45MB)   backbone

来源

  • musetalkV15hf download TMElyralab/MuseTalk
  • sd-vaehf download stabilityai/sd-vae-ft-mse
  • dwposehf download yzd-v/DWPose
  • whisperhf download openai/whisper-tiny
  • syncnethf download ByteDance/LatentSync
  • face-parse-bisent → gdown (Google Drive) + pytorch.org

四、LatentSync 环境

激活conda activate latentsync

关键包版本

版本
torch 2.12.0.dev20260324+cu128nightly
diffusers 0.32.2

已下载权重

位置:~/work/LatentSync/checkpoints/(总计 5.4GB

checkpoints/
├── latentsync_unet.pt             (4.8GB)  主扩散模型
├── whisper/
│   └── tiny.pt                    (73MB)   Whisper 音频编码器
└── auxiliary/
    └── models/buffalo_l/          (约330MB) insightface 人脸模型
        ├── 1k3d68.onnx            (137MB)
        ├── 2d106det.onnx          (4.8MB)
        ├── det_10g.onnx           (17MB)
        ├── genderage.onnx         (1.3MB)
        └── w600k_r50.onnx         (167MB)

来源

  • latentsync_unet.pthf download ByteDance/LatentSync-1.6
  • whisper/tiny.pthf download ByteDance/LatentSync-1.6
  • buffalo_l → 运行时自动下载(insightface

五、磁盘占用汇总

类别 路径 大小
LivePortrait 权重 work/LivePortrait/pretrained_weights 1.2GB
LatentSync 权重 work/LatentSync/checkpoints 5.4GB
MuseTalk 权重 work/MuseTalk/models 5.5GB
模型合计 ~12GB
LivePortrait conda 环境 envs/LivePortrait 8.6GB
MuseTalk conda 环境 envs/MuseTalk 11GB
latentsync conda 环境 envs/latentsync 9.2GB
环境合计 ~29GB

六、产品新环境建议

如果要为产品重建一套干净环境,最小必要集合是:

# 一个统一环境(避免两套 conda env 切换)
conda create -n digital-human python=3.10
conda activate digital-human
conda install -y -c conda-forge ffmpeg

# PyTorch nightlyRTX 5090 必须)
pip install --pre torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/nightly/cu128

# LivePortrait 依赖
pip install -r ~/work/LivePortrait/requirements.txt

# MuseTalk 依赖(含 MMLab 完整链)
pip install -r ~/work/MuseTalk/requirements.txt
pip install --no-build-isolation chumpy
pip install mmengine
MMCV_WITH_OPS=1 pip install mmcv==2.1.0 --no-build-isolation
pip install "mmdet>=3.3.0" "mmpose>=1.3.0"
pip install xtcocotools json-tricks munkres
pip install "transformers>=4.45.0"

可复用的模型权重(无需重新下载):

  • LivePortrait~/work/LivePortrait/pretrained_weights/ — 直接用
  • MuseTalk~/work/MuseTalk/models/ — 直接用
  • LatentSync(如需):~/work/LatentSync/checkpoints/ — 直接用