ComfyUI Dependency Conflicts: Fix Without Reinstalling

A dependency conflict is when one plugin, package, or wheel changes a shared dependency (PyTorch, NumPy, OpenCV, a Hugging Face package) in a way that breaks another plugin, the runtime, or your workflows. Do not reinstall first. Identify which plugin import or package change broke the environment, then repair the smallest possible layer.

If your exact pip check line is sensevoice-onnx requires setuptools<=65.0, use the dedicated SenseVoice-ONNX setuptools conflict guide.

First, confirm the layer

Many "dependency conflict" sessions are really one layer earlier or later than a package conflict. Route before running install commands:

Only continue here when the failure is in the shared Python package stack.

Frequent conflict signatures

Users often search the exact package name from pip check. A pip check line is a clue, not automatically a command — first confirm whether ComfyUI fails, which plugin failed to import, and whether the workflow you need actually uses that package.

Before you install anything else

1. Record the exact error from the startup terminal.
2. Record which plugin or requirement you installed last.
3. Confirm which Python actually launches ComfyUI.
4. Decide whether the failure is a core runtime break, a plugin import break, or a harmless declaration warning.

Step 1: Diagnose

Check for import failures in the startup log:

IMPORT FAILED: ComfyUI-ExampleNode
ModuleNotFoundError: No module named 'somepackage'

Run pip check in the same Python that starts ComfyUI:

Example output:

some-plugin 1.0 requires numpy<2, but you have numpy 2.4.4
another-package 2.0 requires pillow<11, but you have pillow 12.2.0

Check Manager's failure info: if ComfyUI is running, visit http://127.0.0.1:8188/v2/customnode/import_fail_info_bulk for a JSON list of import failures.

Step 2: Classify before installing anything

• Hard blockers (fix first): torch mismatch or CPU-only Torch (see Torch CUDA repair); a missing package a needed plugin directly imports; a native wheel failure like DLL load failed while importing cv2; two required plugins needing non-overlapping versions.
• Soft drift (monitor): pip check reports a declaration mismatch but ComfyUI starts, the plugin imports, and the workflow runs; a package only used by a feature you never use; an optional acceleration backend missing where the node has a PyTorch fallback. Rule of thumb: if ComfyUI starts, the plugin registers, and your workflow runs, do not "fix" the warning just to make pip check look pretty.
• Declaration errors: typos (libros vs librosa), broad pins that downgrade core packages, optional packages listed as mandatory, or Linux-correct pins that are wrong on Windows. Document the workaround; do not edit random plugin source as the first response.

Step 3: Fix

Protect your core runtime. These must not be changed accidentally by a plugin install: torch/torchvision/torchaudio, numpy, pillow, opencv-python/opencv-python-headless, transformers/diffusers/huggingface-hub. Preview what an install will change first:

python -m pip install -r requirements.txt --dry-run

If the dry run wants to replace Torch, NumPy, Pillow, OpenCV, or the Hugging Face stack, stop and decide whether that plugin really needs those changes.

Use the correct Python.

Use minimal fixes. Do not broad-upgrade a shared stack (python -m pip install -U transformers diffusers huggingface-hub) just because one plugin failed. Prefer the narrow fix tied to the exact failing import, or the plugin author's tested compatibility table.

Use the correct package name — some errors are package-name traps:

Use pre-built wheels for native packages (nunchaku, insightface, some attention kernels): match the Python tag (cp312), PyTorch version if tied to Torch, CUDA version if the package includes CUDA kernels, and Windows architecture (win_amd64).

What usually makes conflicts worse

• reinstalling random packages from old Reddit or Discord messages
• using plain pip install against the wrong Python
• letting a plugin downgrade Torch without noticing
• upgrading numpy, opencv, transformers, or diffusers as a group without checking impact
• fixing a Windows Triton error with the normal upstream triton package
• treating every pip check warning like a production outage

Common conflict recipes

• Hugging Face stack drift — diffusers, transformers, huggingface-hub, and accelerate need a compatible set. Check the plugin README or error log, then install the smallest compatible set.
• NumPy 2.x vs 1.x — do not globally downgrade NumPy from an old guide. Confirm a real NumPy error from the plugin you need first; if it truly requires 1.x, isolate that workflow.
• OpenCV conflicts — keep one flavor; opencv-python-headless is the safer default. See OpenCV (cv2) Missing.
• Nunchaku wheel mismatch — use Nunchaku's install workflow or a wheel matching your exact environment; fix the ComfyUI-nunchaku import before installing a package named after a node. See Nunchaku Missing.
• MMCV build/wheel failure — decide whether the plugin needs full MMCV CUDA ops or only mmcv-lite. See MMCV Install Failed.
• ONNX Runtime GPU mismatch — CPU onnxruntime is enough for many pose/face helpers; only install GPU ONNX Runtime when you can match the CUDA/cuDNN family. See ONNX / ONNXRuntime Missing.

Recovery: when your environment is broken

1. Record the current Python path, ComfyUI version, PyTorch/CUDA versions, and the last plugin installed.
2. Restore Torch using the Torch CUDA repair guide.
3. Repair ComfyUI's own requirements in the same environment.
4. Restart and confirm /system_stats works.
5. Add plugin repairs one at a time, checking IMPORT FAILED after each change.

For the official Windows portable package, run commands from the extracted portable root, not from inside the ComfyUI subfolder. If PyTorch and CUDA still work but core packages are missing, use Repair Broken ComfyUI Portable Dependencies before broad commands that may replace the GPU stack.

Before marking the repair finished, verify the same ComfyUI process:

If verification exposes a new missing model, switch to the model-placement path instead of changing more packages.

When to split environments

Consider separate environments when one old plugin and one new plugin require non-overlapping versions, one workflow needs legacy NumPy/OpenCV pins while another needs newer model stacks, or a video/acceleration plugin keeps forcing core runtime changes. If the environment is too damaged to recover, keep a copy of workflows, models, and custom_nodes, re-extract from the Portable Package or Desktop installer, then reinstall plugins one at a time (see Custom Nodes).

FAQ

Should I reinstall ComfyUI right away?

Usually no. First identify whether the break is in Torch, a plugin import, a native wheel, or only a version declaration mismatch.

What command should I run first?

Run pip check in the same Python that launches ComfyUI, then compare it with the actual IMPORT FAILED lines in the startup log.

Why did one custom node break unrelated workflows?

Many plugins share the same Python environment and package stack. A change meant for one node can replace core packages used by many others.

Related Guides

• Repair Broken ComfyUI Portable Dependencies Without Reinstalling Torch
• ComfyUI No Module Named 'sageattention'
• ComfyUI Triton Missing or Unavailable
• ComfyUI Nunchaku Missing
• ComfyUI MMCV Install Failed
• ComfyUI Torch Not Compiled With CUDA Enabled
• Plugin Management
• Troubleshooting Decision Tree
• Common Issues

Source References

• Official ComfyUI custom node troubleshooting
• Official ComfyUI model troubleshooting
• Official ComfyUI portable Windows guide
• Official ComfyUI update guide