ComfyUI Common Issues and Fast Fixes

This is the routing page for common ComfyUI failures. Pick the closest match and jump to the dedicated guide instead of repeating command-line repairs.

Pick the right fix first

Installation issues

"CUDA is not available" or "Torch not compiled with CUDA"

PyTorch was installed without CUDA support, or the CUDA version doesn't match your driver.

1. Check your NVIDIA driver: nvidia-smi
2. Confirm the Python that launches ComfyUI has a CUDA-enabled Torch build:

python -c "import torch; print(torch.__version__); print(torch.version.cuda); print(torch.cuda.is_available())"

3. For the portable package, download the NVIDIA version and run run_nvidia_gpu.bat.
4. If it prints False, follow the Torch not compiled with CUDA enabled guide before installing more plugins.

ComfyUI Desktop shows "Unsupported device"

Desktop on Windows requires an NVIDIA GPU with CUDA. Use the Portable Package or Manual Install for AMD/Intel GPUs.

Installer or app blocked by antivirus

Add ComfyUI's install directory to your antivirus exception list. For Windows Defender: Settings → Privacy & Security → Virus & Threat Protection → Manage Settings → Exclusions, then re-download.

7-Zip extraction fails

1. Right-click the .7z → Properties → check Unblock → Apply
2. Extract to a short path like D:\ComfyUI
3. Use 7-Zip, not Windows built-in zip.

"Permission denied" or package install failures

1. Never run ComfyUI as Administrator — it causes Python package permission conflicts.
2. Do not install to C:\Program Files, C:\Windows, or C:\ root.
3. Use a non-system path like D:\ComfyUI or C:\Users\YourName\ComfyUI.

Windows long path limit

Deep custom-node folders can exceed the 260-character limit. Enable long paths:

1. Win + R, type regedit, Enter
2. Go to HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\FileSystem
3. Set LongPathsEnabled to 1 and restart.

Or in PowerShell (as Administrator):

New-ItemProperty -Path "HKLM:\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1 -PropertyType DWORD -Force

Runtime issues

CUDA out of memory

Your GPU lacks VRAM for the model or resolution. See System Requirements. Try in order: close other GPU apps; reduce resolution (512×512 vs 1024×1024); add --lowvram; use GGUF quantized models; for video, reduce frame count and resolution. If the log says MemoryError, DefaultCPUAllocator, or MPS backend out of memory, use the full Out of Memory guide.

Red nodes in workflow

The workflow uses custom nodes you haven't installed. Install ComfyUI Manager, click Install Missing Custom Nodes, and restart.

"No checkpoint found" / empty model dropdown

Place the .safetensors file in ComfyUI/models/checkpoints/, then click Refresh or restart. For LoRA, VAE, ControlNet, Flux, GGUF, and text encoder folders, see Where to Put Safetensors.

Browser shows blank page or title only

Use the latest Google Chrome. If the terminal mentions comfyui-frontend-package, see ComfyUI Frontend Package.

ComfyUI keeps reconnecting

The browser lost its live connection. Check whether the terminal is still running, test with custom nodes disabled, and open from http://127.0.0.1:8188 before LAN or proxy. See Reconnecting Error.

Failed to fetch server logs

The browser's log request failed — often after a crash, firewall block, or broken frontend extension. Find the real terminal log first. See Failed to Fetch Server Logs.

"Prompt has no outputs"

The workflow has no output node, or it is muted/disconnected. Connect a Save Image or Preview Image node; press M to unmute greyed nodes; install red (missing) nodes via Manager. See Prompt Has No Outputs.

Generation is extremely slow

ComfyUI may be on CPU. Check the console for your GPU name at startup; for portable, run run_nvidia_gpu.bat, not run_cpu.bat. For manual installs:

python -c "import torch; print(torch.cuda.is_available())"

Should print True. If False, follow Torch Missing with the Python that starts ComfyUI. If you have more than one NVIDIA GPU and the wrong card is used, see Multi GPU.

"No module named 'triton'" or Triton unavailable

Triton is an optional acceleration backend. If ComfyUI opens and normal workflows run, this is usually not the root problem. Do not install the normal triton package blindly on Windows — see Triton Missing or Unavailable.

GPU-specific issues

RTX 50 Series (5070 Ti / 5080 / 5090)

See GPU Compatibility for the full driver-CUDA-PyTorch matrix. The RTX 50 series needs CUDA 12.8+ and specific PyTorch builds:

1. Update your NVIDIA driver.
2. Install PyTorch using the matching command from the official PyTorch selector in the Python that starts ComfyUI. Its CUDA-wheel pattern is:

python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cuXXX

Replace cuXXX with the CUDA wheel tag for your supported build. Do not use it as a generic copy-paste command or add a third-party mirror. 3. For portable, use the latest release with the required CUDA support.

Common 50-series issues: SageAttention compilation errors (see SageAttention Missing); Nunchaku plugin failures (see Nunchaku Missing); xformers crashes (use --use-pytorch-cross-attention).

AMD GPUs on Windows

See GPU Compatibility. AMD on Windows uses DirectML, which has limitations: some nodes don't support it, performance is lower, and you must pass --directml in a real launch command (never add a GPU flag to run_cpu.bat).

Network issues

Hugging Face model download fails

Root cause may be network routing, proxy, gated access, an invalid token, offline mode, or a wrong path. Test the exact Hugging Face URL in a browser and terminal; if the browser works but ComfyUI fails, check whether the ComfyUI process has the same proxy and HF_TOKEN. See HuggingFace HttpRequestException.

pip install hangs or times out

Network, firewall, or proxy is blocking package downloads. If behind a proxy:

pip install --proxy http://your-proxy:port -r requirements.txt

In ComfyUI Desktop, change the mirror settings in the setup wizard. If your organization requires a regional mirror, use its documented and trusted endpoint as a network policy; do not add an unfamiliar mirror to a repair command blindly.

Git clone fails

GitHub is blocked or rate-limited in your region. Use a mirror or download the ZIP from the GitHub release page.

Not finding your issue?

Try the Troubleshooting Decision Tree for a systematic approach. For complex multi-plugin environments, see Dependency Conflicts and Workflow Environment Setup.

Source References

• ComfyUI troubleshooting overview
• ComfyUI custom node troubleshooting guide
• ComfyUI system requirements