This is a numpy/mediapipe version conflict. Mediapipe requires numpy 1.26.4 but other packages may have upgraded it.

v0.3.1 Fix (recommended):
The installer CLI handles this automatically. Activate your venv from the Install page and run:

cd StreamDiffusion-installer
python -m sd_installer repair

The repair command pins numpy, reinstalls mediapipe, and fixes opencv in one step.

v0.3.0 Manual Fix:

1. Save the script below as Fix_Mediapipe_Numpy.bat (or download from Discord)


2. Place it in your StreamDiffusionTD base folder (where the venv folder is)


3. Double-click to run

@echo off
echo.
echo ========================================
echo  StreamDiffusionTD - Mediapipe/NumPy Fix
echo ========================================
echo.

cd /d "%~dp0"
if not exist "venv" (
    echo ERROR: venv folder not found in %CD%
    pause
    exit /b 1
)
call "venv\Scripts\activate.bat"
if "%VIRTUAL_ENV%" == "" (
    echo ERROR: Failed to activate venv
    pause
    exit /b 1
)
echo Activated: %VIRTUAL_ENV%
echo.
echo [1/6] Removing conflicting packages...
pip uninstall -y opencv-python-headless mediapipe opencv-python opencv-contrib-python
echo.
echo [2/6] Locking numpy FIRST (prevents drift)...
pip install --no-cache-dir "numpy==1.26.4" --force-reinstall
echo.
echo [3/6] Installing mediapipe with --no-deps (prevents numpy upgrade)...
pip install --no-cache-dir --no-deps "mediapipe==0.10.18"
echo Installing mediapipe dependencies manually...
pip install --no-cache-dir absl-py attrs flatbuffers matplotlib protobuf sounddevice
echo.
echo [4/6] Installing opencv-python with --no-deps...
pip install --no-cache-dir --no-deps "opencv-python==4.8.1.78"
echo.
echo [5/6] Fixing insightface (if present)...
pip show insightface >nul 2>&1 && (
    pip uninstall -y insightface
    pip install --no-cache-dir --no-deps insightface
    pip install --no-cache-dir --no-deps onnxruntime prettytable
)
pip uninstall -y opencv-python-headless opencv-contrib-python 2>nul
echo.
echo [6/6] Fixing onnx version (onnx_graphsurgeon compatibility)...
pip install --no-cache-dir "onnx<1.20.0"
echo.
echo Re-locking numpy (final safety check)...
pip install --no-cache-dir "numpy==1.26.4" --force-reinstall
echo.
echo ========================================
echo Verification
echo ========================================
python -c "import numpy; print(f'numpy {numpy.__version__}')"
python -c "import cv2; print(f'opencv {cv2.__version__}')"
python -c "import mediapipe as mp; mp.solutions.drawing_utils; print('mediapipe OK')" || goto :error
python -c "import torch; print(f'torch {torch.__version__} CUDA:{torch.cuda.is_available()}')"
python -c "from streamdiffusion.config import load_config; print('StreamDiffusion OK')" || echo StreamDiffusion FAILED
echo.
echo ========================================
echo FIX COMPLETE
echo ========================================
echo.
pause

The script locks numpy to 1.26.4, reinstalls mediapipe 0.10.18 with --no-deps to prevent version drift, and fixes opencv versions.