ComfyUI Startup Failed? How to Diagnose and Recover Faster

When ComfyUI won't start or fails before Starting server, find the first real startup error, then route to the narrow repair page. Most startup failures happen because the environment changed after install — a plugin modified a critical package, a binary package no longer matches the runtime, Torch or CUDA drifted, or a helper/bootstrap file was removed or quarantined. The real question is not "how do I launch ComfyUI?" but "what changed between the last working state and now?"

Route by the first error

Step 1: Classify the startup failure

• Category A — Python import failure during startup (ModuleNotFoundError: No module named 'sqlalchemy', comfyui-frontend-package is not installed, No module named 'cv2'). A package is missing or broken. If it is imported by ComfyUI's own startup path, use the portable dependency repair runbook rather than treating it as one custom node.
• Category B — Core runtime drift (Torch not compiled with CUDA, CUDA is not available, AttributeError: module 'torch' has no attribute '...'). Torch, CUDA, or another critical dependency changed.
• Category C — Plugin import chain blocks startup (one node fails and startup becomes unstable; many plugins fail after one install; it worked before a plugin or update). See plugin import failures and dependency conflicts.
• Category D — Bootstrap or launcher helper failure (a helper executable is missing, antivirus quarantined a file, a launcher cannot finish booting). Check antivirus quarantine before reinstalling.

Step 2: Fix the narrowest blocker first

When the machine used to work and now does not, this order is safest:

1. Read the first real startup error — the earliest IMPORT FAILED, ModuleNotFoundError, Torch/CUDA/DLL error, or missing helper. Everything after it may be fallout.
2. Decide whether it is core runtime, plugin dependency, or deployment/bootstrap.
3. Use the smallest recovery move for that category:
   • Missing package — install only that package or the plugin requirements that introduced it; do not broad-upgrade everything. For No module named 'sqlalchemy', see the SQLAlchemy startup fix; if fixing one package reveals another core one, switch to the portable dependency repair runbook.
   • Torch/CUDA drift — repair the core runtime first, then retest plugin imports. Common signs: unexpected CPU fallback, CUDA stopped after a custom node install, or xformers/onnxruntime pulled a mismatched build.
   • Missing helper/bootstrap file — check antivirus quarantine, install directory integrity, and launcher helper files. See Bootstrapper Missing.
4. Restart and re-check the new first error.
5. Consider reinstall only when the environment no longer has a stable first blocker — and only after preserving models, workflows, and custom nodes.

When startup failure is really a deployment failure

Some first-run issues get described as "startup failed" even though ComfyUI never reached a stable installed state — the resource package did not finish downloading, ComfyUI-Manager did not clone, or a bootstrap file was blocked. If the launcher fails before your first stable launch, check these before repairing Python packages:

• Deployment Failed: Downloading the Resource Package
• Deployment Failed: Installing ComfyUI-Manager
• Bootstrapper Missing (watchdog_bootstrapper_missing)

Related guides

• Repair Broken ComfyUI Portable Dependencies Without Reinstalling Torch
• ModuleNotFoundError: No module named 'sqlalchemy' in ComfyUI
• How to Fix ComfyUI Plugin Import Failed Errors
• ComfyUI Dependency Conflicts
• ComfyUI Reconnecting Error
• Common Issues
• Troubleshooting Decision Tree

Source References

• ComfyUI troubleshooting overview
• ComfyUI custom node troubleshooting guide
• ComfyUI model troubleshooting guide
• ComfyUI Desktop installation guide