CLI tool for installing and repairing StreamDiffusion environments used by StreamDiffusionTD.
Clone this repo into your StreamDiffusion folder:
cd /path/to/StreamDiffusion
git clone https://github.com/dotsimulate/StreamDiffusion-installer.gitRun from the StreamDiffusion folder or from inside StreamDiffusion-installer:
# Check system readiness
python -m sd_installer check
# Fresh install with CUDA 12.8 (recommended)
python -m sd_installer install --cuda cu128
# Install with specific CUDA version
python -m sd_installer install --cuda cu121
# Verify existing installation
python -m sd_installer verify
# Detailed diagnostics
python -m sd_installer diagnose
# Auto-fix known issues
python -m sd_installer repair
# Generate standalone batch file
python -m sd_installer generate-bat --cuda cu128| CUDA | PyTorch | GPU Support |
|---|---|---|
| cu118 | 2.4.0 | RTX 20/30 series |
| cu121 | 2.4.0 | RTX 30/40 series |
| cu124 | 2.4.0 | RTX 30/40 series |
| cu128 | 2.7.0 | RTX 30/40/50 series (recommended) |
The installer uses the correct dependency sequence:
- numpy locked first - Prevents numpy 2.x conflicts
- PyTorch with correct CUDA - Before any other packages
- setup.py handles versions - Single source of truth for pins
- Post-install fixes - timm, opencv with --no-deps
- Verification - Import tests catch failures immediately
| Error | Cause | Fix |
|---|---|---|
cannot import name 'RotaryEmbedding' |
timm too old | timm>=1.0.24 |
cannot import name 'cached_download' |
huggingface_hub version | huggingface_hub==0.35.0 |
cannot import name 'MT5Tokenizer' |
transformers too new | transformers==4.56.0 |
module 'mediapipe' has no attribute 'solutions' |
mediapipe conflict | --no-deps mediapipe==0.10.21 |
numpy 2.x errors |
numpy upgraded by dependencies | numpy==1.26.4 locked |
At runtime, the expected structure is:
StreamDiffusion/ <- base folder (par.Basefolder in TD)
├── src/
├── setup.py
├── venv/ <- created by installer
├── streamdiffusionTD/ <- TD extension files
└── StreamDiffusion-installer/ <- this repo
└── sd_installer/