Installation Guide (v0.4.0)

Pick your route

The local operator runs on your own NVIDIA GPU with TensorRT and the full feature set (ControlNet, IP Adapter, StreamV2V, live LoRA weight, FX processors) in one session. Already have a v0.3.1 install? Switch to Upgrade above.

Prerequisites Python 3.11.9, Git on PATH, NVIDIA GPU with a CUDA 12.8 driver, Windows 10/11

Python 3.11.9 (required): Download here

  • 3.11.9 is the version this release is built and tested against
  • You do not need to check “Add to PATH” during installation
  • Not compatible: Python 3.12, 3.13. Do not use 3.10 for v0.4.0 (cuda-link ships as a Python 3.11 wheel)

Git: Download here and check “Add to PATH”

GPU: NVIDIA, Windows 10/11. CUDA 12.8 is the only option the operator offers. You need a driver that can run CUDA 12.8. You do not need to pick a CUDA version in the UI anymore.

Install folder: do not name it streamdiffusion (that collides with the Python module). Avoid OneDrive paths and paths with spaces. Good examples: C:/AI/StreamDiffusion_v040/, D:/SD/StreamDiff040/

Installation Steps

  1. Load Operator: Drag the streamdiffusiontd_latest.tox into TouchDesigner and click Sign in. It replaces itself with the operator
  2. Install Page: Go to the Install parameter page
  3. Set Base Folder: Choose a new, empty folder
  4. Step 1, Clone StreamDiffusion: downloads the pinned v0.4.0 repos from the version manifest (SDTD_040_stable / SDTD_040_installer)
  5. CUDA: 12.8. There is no other choice in this release
  6. Step 2, Install StreamDiffusion: builds the venv (PyTorch 2.11.0+cu128, cuda-link, the rest of the stack). Wait for the console to finish
  7. Step 3, Install TensorRT: TensorRT 10.16 via tensorrt_cu12 sub-packages. Required for ControlNet, IP Adapter, and StreamV2V
  8. Restart TouchDesigner after a first-time install so CUDALINK_LIB_PATH is visible. cuda-link is how frames stay on the GPU between TD and the stream

Verify

Activate the venv from the Install page, then:

cd StreamDiffusion-installer
python -m sd_installer verify

If anything fails:

python -m sd_installer repair

For a bug report dump:

python -m sd_installer diagnose

First Run

  1. Models page: Acceleration = tensorrt (the menu is none / tensorrt. xformers is gone)
  2. Click Start Stream
  3. First TensorRT engine build takes a while. Let it finish
  4. Default model: SDXL-Turbo

Suggested first-test config:

  • Model: stabilityai/sd-turbo if you want a lighter first engine
  • Resolution: 512x512
  • IP Adapter: OFF
  • ControlNet: enabled, weight 0
  • Acceleration: TensorRT

The first ControlNet stream will rebuild the ControlNet engine once. Old engine folders are left in place. That is expected.

What v0.4.0 installs PyTorch 2.11 (cu128), TensorRT 10.16, cuda-link 1.12.2, Python 3.11.9
PieceVersion
PyTorch / torchvision2.11.0 / 0.26.0 (cu128)
TensorRT10.16.1.11 (tensorrt_cu12 sub-packages)
CUDA selector12.8 only
cuda-link1.12.2 (prebuilt wheel, no MSVC needed)
Python3.11.9

TensorRT is required for ControlNet, IP Adapter, and StreamV2V. acceleration: none is still broken upstream.

Common Installation Issues

Slow install / pypi.ngc.nvidia.com warnings

If logs retry pypi.ngc.nvidia.com or the install crawls, that is a leftover machine-wide pip config from pre-v0.3.0. Updating StreamDiffusionTD does not remove it. See this FAQ.

“polygraphy” Module Missing

Error: ModuleNotFoundError: No module named 'polygraphy' Fix: You skipped Step 3. Click Install TensorRT.

Python Version Mismatch

Fix: python --version must be 3.11.x. Uninstall 3.12/3.13 or fix PATH.

Folder Naming Conflict

Error: Import failed: no module named 'streamdiffusion.config' Fix: Do not name the installation folder streamdiffusion.

Acceleration Has Failed

Fix: Delete StreamDiffusion/engines/td/ and restart. For v0.4.0 betas this is required anyway.

flash-attn Crash

Fix: pip uninstall flash-attn in the StreamDiffusion venv. It is not used.

OpenCV INTER_AREA

Error: AttributeError: module 'cv2' has no attribute 'INTER_AREA' Fix: repair does not handle this. See the FAQ.

Troubleshooting

For more errors, use the FAQ search (those URLs actually deep-link). The Troubleshooting page is the long list.

Paste python -m sd_installer diagnose output in Discord #troubleshoot-central and tag @dotsimulate.