The first fully documented, Windows-native ComfyUI setup for NVIDIA GeForce RTX 5090/5080/5070 (Blackwell architecture, sm_120) with CUDA 13.0.
RTX 50-series GPUs (Blackwell, Compute Capability sm_120) are not supported by PyTorch stable releases as of early 2026. Running ComfyUI on these GPUs requires specific versions and workarounds that are not documented anywhere else in a single, reproducible package.
| Feature | Details |
|---|---|
| GPU Architecture | NVIDIA Blackwell (sm_120) -- RTX 5090 / 5080 / 5070 |
| CUDA Version | 13.0 (cu130) -- the latest CUDA runtime |
| PyTorch | Nightly cu130 build (not stable, not cu128) |
| Python | 3.13 (latest) |
| Triton | triton-windows fork (official Triton is Linux-only) |
| xformers | Deliberately excluded (causes PyTorch downgrade) |
| Custom Nodes | 28 verified nodes including video & music generation |
| Platform | Windows Native (no WSL2, no Docker required) |
-
Blackwell + Windows Native + CUDA 13.0 -- One of the world's first documented setups that runs ComfyUI on Blackwell GPUs entirely on Windows without WSL2 or Docker.
-
No xformers Architecture -- Pioneers the use of
triton-windows+torch.compileas a replacement for xformers, which is incompatible with Blackwell nightly builds. -
28 Custom Nodes Verified -- All nodes have been individually tested on RTX 5090 (32GB VRAM), including cutting-edge video generation (Wan 2.1, LTX-Video, HunyuanVideo 1.5, Kandinsky 5.0, LongCat-Video) and music generation (ACE-Step, HeartMuLa).
-
5 Image-to-Video Pipelines Verified -- Complete I2V workflow validation on Blackwell hardware.
-
Windows-specific Fixes Documented -- Path separator issues, SageAttention fallback to SDPA, Triton compilation workarounds, and more.
| Component | Version |
|---|---|
| OS | Windows 11 Pro (Build 26200) |
| GPU | NVIDIA GeForce RTX 5090 (32GB VRAM) |
| CPU | Intel Core Ultra 9 285K (Arrow Lake) |
| RAM | 64GB DDR5-5600 |
| NVIDIA Driver | 591.55 |
| CUDA | 13.0 |
| Python | 3.13 |
| PyTorch | 2.9.1+cu130 |
| triton-windows | 3.6.0 |
- Windows 10/11 (64-bit)
- NVIDIA GeForce RTX 5090 / 5080 / 5070 (or any Blackwell GPU)
- NVIDIA Driver 580+ (Download)
- Git (Download)
- 7-Zip (Download) for extracting the portable build
1. Download or clone this repository
2. Double-click setup.bat
3. Wait for the installation to complete (~20 minutes)
4. Double-click run.bat to start ComfyUI
5. Open http://localhost:8188 in your browser
# Clone this repository
git clone https://github.com/hiroki-abe-58/ComfyUI-GeForce-Blackwell.git
cd ComfyUI-GeForce-Blackwell
# Run the setup script
powershell -ExecutionPolicy Bypass -File setup.ps1
# Start ComfyUI
.\run.bat# After setup, verify everything is working
python verify_env.pyExpected output:
============================================================
Blackwell (sm_120) Environment Verification
============================================================
[OK] Python
[OK] NVIDIA Driver
[OK] PyTorch
[OK] Triton
[OK] Core Packages
[OK] torch.compile
Environment is ready for Blackwell GPU!
ComfyUI-GeForce-Blackwell/
├── setup.bat # One-click setup (double-click)
├── setup.ps1 # Full setup script (PowerShell)
├── run.bat # Start ComfyUI
├── update.bat # Update everything
├── update.ps1 # Update script (PowerShell)
├── verify_env.py # Environment verification
├── requirements-core.txt # Core dependencies (torch excluded)
├── custom-nodes.json # 28 verified custom nodes
├── configs/
│ └── extra_model_paths.yaml.example
└── scripts/
├── install_custom_nodes.ps1 # Batch node installer
└── fix_windows_compat.py # Workflow compatibility fixer
These rules must be followed. Violating any of them will break your environment.
| # | Rule | Why | What Happens If Violated |
|---|---|---|---|
| 1 | Use PyTorch nightly cu130 | Stable builds don't include sm_120 kernels | RuntimeError: sm_120 is not compatible |
| 2 | Never install xformers | It force-downgrades PyTorch to stable | Everything stops working |
| 3 | Exclude torch from requirements.txt | pip will overwrite nightly with stable | Silent version downgrade |
| 4 | Verify after adding custom nodes | Node dependencies may pull in stable torch | Previously working setup breaks |
| 5 | Clear proxy environment variables | System proxies block pip/git connections | Installation failures |
- ComfyUI-WanVideoWrapper -- Wan 2.1 video generation
- ComfyUI-LTXVideo -- LTX-Video generation
- comfyui-videohelpersuite -- Video combine & helpers
- comfyui_longcat_image -- LongCat video processing
- ComfyUI-AceMusic -- ACE-Step music generation
- ComfyUI-HeartMuLa -- HeartMuLa music generation
- ComfyUI-MelBandRoFormer -- Audio source separation
- ComfyUI_FL-HeartMuLa -- HeartMuLa FL variant
- ComfyUI-Step1X-Edit -- Step1X image editing
- comfyui-impact-pack -- Detection & inpainting
- comfyui_controlnet_aux -- ControlNet preprocessors
- comfyui_layerstyle -- Layer compositing
- comfyui-depthanythingv2 -- Depth estimation
- ComfyUI-Manager -- Node management
- comfyui-kjnodes -- KJ utility nodes
- comfyui_essentials -- Essential utilities
- was-node-suite-comfyui -- WAS Node Suite
- ComfyUI_Comfyroll_CustomNodes -- Comfyroll collection
- rgthree-comfy -- Power tools
- efficiency-nodes-comfyui -- Efficient sampling
- masquerade-nodes-comfyui -- Mask manipulation
- comfyui-easy-use -- Simplified workflows
- comfyui-custom-scripts -- UI enhancements
- ComfyUI-Crystools -- System monitoring
- comfyui-quadmoons-nodes -- Quadmoons utilities
- ComfyUI_SimpleTranslator -- Text translation
- ComfyUI_kkTranslator_nodes -- Translation nodes
- ComfyUI-SKBundle -- SK bundle utilities
| Model | Parameters | FP8 Size | Performance on 32GB VRAM |
|---|---|---|---|
| HunyuanVideo 1.5 I2V | 8.3B | ~16GB | Smooth -- recommended |
| Kandinsky 5.0 Lite I2V | 2B | ~4GB | Very smooth |
| LTX-2 I2V | 19B | ~25GB | Works in FP8 |
| LongCat-Video TI2V | 13.6B | ~14.5GB | Works with adjustments |
| Kandinsky 5.0 Pro I2V | 19B | ~40GB | Requires CPU offload, slow |
Linux workflows use forward slashes (/) in model paths. Windows requires backslashes (\).
Automatic fix:
python scripts/fix_windows_compat.py your_workflow.jsonSageAttention is difficult to build on Windows. Use SDPA (PyTorch native) instead:
- Change
attention_modefromsageattntosdpain your workflow - Or use
fix_windows_compat.pyto fix automatically
triton-windows may not work perfectly in all scenarios. If you see Triton compilation errors:
- Disconnect
compile_argsfrom model loaders in your workflow - RTX 5090 is fast enough without torch.compile for most workflows
After installing any package, always verify:
python -c "import torch; print(torch.__version__)"If it no longer contains cu130, reinstall:
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130Double-click update.bat or run:
powershell -ExecutionPolicy Bypass -File update.ps1This will update ComfyUI core, PyTorch, and all custom nodes while preserving Blackwell compatibility.
Q: Do I need WSL2 or Docker? A: No. This setup runs entirely on Windows native. WSL2/Docker are alternatives but not required.
Q: Can I use xformers? A: No. xformers forces a downgrade to PyTorch stable, which doesn't support sm_120. Use triton-windows + torch.compile instead.
Q: Will this work on RTX 4090 (Ada Lovelace)? A: Yes. cu130 is backward-compatible. However, RTX 4090 users can also use stable PyTorch builds.
Q: Where do I put models?
A: After setup, place models in ComfyUI_windows_portable/ComfyUI/models/. See configs/extra_model_paths.yaml.example for shared model storage.
Q: When will PyTorch stable support Blackwell? A: Expected in a future release (possibly PyTorch 2.12+). Until then, nightly cu130 is required.
Pull requests are welcome. If you've verified additional custom nodes or workflows on Blackwell hardware, please submit a PR to update custom-nodes.json.
MIT License -- see LICENSE
- ComfyUI by Comfy-Org
- triton-windows by woct0rdho
- PyTorch nightly team for sm_120 support
- All custom node developers listed in
custom-nodes.json
