Fix the dependency errors that block custom nodes from loading
#How to tell it's a dependency problem
If ComfyUI Manager's "Install Missing Custom Nodes" doesn't fix a red/broken node, or the node loads but throws an error the first time you queue a prompt, the custom node's code is installed but one of its Python or system dependencies isn't. Check the ComfyUI console window — the real error is almost always a few lines above the line ComfyUI prints in the UI.
#"ModuleNotFoundError: No module named 'X'"
Cause: The custom node's requirements.txt never ran, or ran against the wrong Python environment.
Fix:
If that still fails, you likely have two Pythons on your system and installed the node while the wrong one was active. Confirm with:
The path printed should point inside your ComfyUI venv folder. If it doesn't, re-activate the venv and reinstall.
#FFmpeg not found (video export, VideoHelperSuite)
Cause: VideoHelperSuite, AnimateDiff exports, and most LTX Video workflows call the ffmpeg binary directly — it's a system dependency, not a Python package, so pip install can't fix it.
Fix (Windows):
Close and reopen your terminal, then verify:
If winget isn't available, download a build from ffmpeg.org, unzip it, and add the bin folder to your system PATH.
Fix (Linux):
#Face/detection nodes fail on import (insightface, onnxruntime, mediapipe)
Cause: These packages need compiled binaries that match your exact Python version and OS. A pip install that "succeeds" can still fail to import if the wheel doesn't match.
Fix:
On Windows, insightface sometimes needs the Microsoft C++ Build Tools if no precompiled wheel exists for your Python version — install "Desktop development with C++" from the Visual Studio Installer, then retry.
If you have a CUDA GPU, make sure you installed onnxruntime-gpu, not plain onnxruntime — the CPU-only package installs silently and just makes face nodes very slow instead of erroring.
#"CUDA driver version is insufficient" or torch/CUDA mismatch
Cause: A custom node pinned a torch version that doesn't match the CUDA build you installed ComfyUI with, and reinstalling it silently swapped your PyTorch build.
Fix: Reinstall the CUDA-matched PyTorch build explicitly, after any custom node installs:
Confirm it stuck:
torch.cuda.is_available() must print True. If it prints False after this, your NVIDIA driver is older than the CUDA build requires — update the driver first.
#Two custom nodes conflict on the same package version
Cause: Node A needs numpy<2.0, Node B needs numpy>=2.0. Whichever installed last wins, and the other silently breaks.
Fix: There's no universal fix here — check each node's GitHub issues for the specific version pin, and prefer forks/nodes that are actively maintained. As a last resort, keep the two nodes in separate ComfyUI installs rather than fighting version pins in one environment.
#Quick diagnostic checklist
| Symptom | Likely cause | Guide |
|---|---|---|
| Red node, "Install Missing Nodes" doesn't fix it | requirements.txt didn't run | This guide, above |
| Video export nodes error, images work fine | FFmpeg not on PATH | This guide, above |
| Face/detailer nodes crash on first run | onnxruntime/insightface wheel mismatch | This guide, above |
Everything worked yesterday, now torch.cuda.is_available() is False | A node install swapped your torch build | This guide, above |
| Error happens loading the workflow, not a specific node | Wrong model type, not a dependency | Workflow Errors |
| "CUDA out of memory" during generation | VRAM, not a missing package | GPU Errors |