comfyui下载地址：https://docs.comfy.org/installation/comfyui_portable_windows

运行update下的update_comfyui_and_python_dependencies.bat

安装Manager

1. 进入你的新目录

F:\ComfyUI_windows_portable_nvidia_cu126_new\ComfyUI_windows_portable

2. 访问下面这个官方 bat 文件链接,保存install-manager-for-portable-version.bat，和 run_nvidia_gpu.bat 同级，运行

https://github.com/ltdrdata/ComfyUI-Manager/raw/main/scripts/install-manager-for-portable-version.bat

安装需要的 Custom Nodes

Manager安装ComfyUI-VideoHelperSuite：v1.5.10

手动安装ComfyUI-ReActor：

1. 下载insightface-0.7.3-cp312-cp312-win_amd64.whl，下载后把 .whl 文件放到 ComfyUI_windows_portable 根目录

https://github.com/Gourieff/Assets/raw/main/Insightface/insightface-0.7.3-cp312-cp312-win_amd64.whl

2. 然后执行：

.\python_embeded\python.exe -m pip install insightface-0.7.3-cp312-cp312-win_amd64.whl

3. 安装ReActor

cd ComfyUI\custom_nodes
git clone https://github.com/Gourieff/ComfyUI-ReActor
cd ComfyUI-ReActor
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt

Manager安装Wav2Lip Node for ComfyUI

安装 CPU-only 版 torchcodec

cd /d F:\ComfyUI_windows_portable_nvidia_cu126_new\ComfyUI_windows_portable

:: 2. 安装 CPU-only 版 torchcodec（这是目前 Windows 上最稳的方式）
.\python_embeded\python.exe -m pip install torchcodec --index-url https://download.pytorch.org/whl/cpu

修改 Wav2Lip 节点源码，彻底绕过 torchcodec

1. 打开文件：

F:\ComfyUI_windows_portable_nvidia_cu126_new\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_wav2lip\wav2lip.py

2. 找到 torchaudio.save 调用处（大约在 process 方法第 160~170 行附近）：

torchaudio.save(temp_audio_path, waveform_tensor, sample_rate)

3. 替换为以下代码（强制使用 soundfile 后端，稳定且兼容）：

# === 企业级修复：强制使用 soundfile 后端，彻底绕过 torchcodec ===
import soundfile as sf
import torch

# 如果 waveform 是 torch.Tensor，先转 numpy
if isinstance(waveform_tensor, torch.Tensor):
    waveform_np = waveform_tensor.squeeze().cpu().numpy()
    if len(waveform_np.shape) == 1:
        waveform_np = waveform_np.reshape(1, -1)  # soundfile 需要 (channels, samples) 或 (samples,)
else:
    waveform_np = waveform_tensor

sf.write(temp_audio_path, waveform_np.T if waveform_np.shape[0] < waveform_np.shape[1] else waveform_np, sample_rate)
# =================================================================

4. 彻底修复 ReActor + ReActorRestoreFace（追求高清质量）

cd /d F:\ComfyUI_windows_portable_nvidia_cu126_new\ComfyUI_windows_portable

:: 1. 先清理冲突库（之前 soxr、scipy 等问题）
.\python_embeded\python.exe -m pip uninstall -y soxr torchcodec insightface

:: 2. 安装最稳版本
.\python_embeded\python.exe -m pip install soxr==0.5.0.post1 --force-reinstall
.\python_embeded\python.exe -m pip install soundfile --force-reinstall

:: 3. 安装 insightface（ReActor 核心）
.\python_embeded\python.exe -m pip install https://github.com/Gourieff/Assets/raw/main/Insightface/insightface-0.7.3-cp312-cp312-win_amd64.whl

:: 4. 重新安装 ReActor（使用稳定分支）
cd ComfyUI\custom_nodes
rd /s /q ComfyUI-ReActor 2>nul
git clone https://github.com/Gourieff/ComfyUI-ReActor
cd ComfyUI-ReActor
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt