ComfyUI Queue Errors: Fix Stuck Prompts, Missing Nodes and Validation Failures

Diagnostic Guide Start with evidence, not random reinstalls

When ComfyUI stops at “queued,” the visible symptom rarely identifies the cause. The failure may happen before execution during workflow validation, during model loading, inside a custom node, or after execution when the browser loses its live connection. The fastest fix is to locate the failing layer.

First rule: save the workflow JSON and copy the complete error text before changing the installation. A screenshot of the red node is useful; the full log is better.

1. Identify the Failure Stage

Symptom	Likely layer	First check
Prompt rejected immediately	Validation	Missing inputs, models, or node types
Queue never starts	Server/connection	Backend log and browser connection
Stops on one node	Node execution	Full traceback and node ID
Process closes	Memory/dependency	Terminal log, RAM, VRAM, Python packages

2. Fix Validation Errors Before Touching CUDA

The ComfyUI server validates the complete submitted workflow before adding it to the execution queue. Check red nodes for empty required inputs, model names that no longer exist, invalid widget values, and links with incompatible data types. If an API client receives node_errors, log that object rather than retrying the same payload.

if response.status_code != 200:
    print(response.text)
    raise RuntimeError("ComfyUI rejected the workflow")

For API architecture and batching, see Automating ComfyUI with the API.

3. Resolve Missing Custom Nodes Safely

Use ComfyUI Manager's missing-node view and inspect the repository before installation. Confirm the node pack is maintained and compatible with your frontend. If the node is not in the registry, do not download a similarly named random repository.

If the interface breaks after an update, start once with all custom nodes disabled. When the issue disappears, enable half of the nodes, test, and repeat. This binary-search method finds a conflict much faster than reinstalling everything.

Dependency warning: install dependencies into ComfyUI's own Python environment. Mixing portable Python, system Python, and another virtual environment creates errors that look unrelated.

4. Separate a Stuck Queue From a Disconnected Browser

ComfyUI uses HTTP to submit the prompt and a WebSocket connection for live progress. A render can continue even when the browser stops updating. Check the backend queue and output directory before interrupting a long job. For remote setups, inspect reverse-proxy WebSocket support and timeouts.

5. Diagnose CUDA and Memory Failures

1. Record the exact node where allocation fails.
2. Test a smaller resolution and batch size of one.
3. Close other GPU-heavy applications.
4. Use the correct low-VRAM approach for the model.
5. Restart only after saving the error and workflow.

The FLUX workflow optimization guide covers memory pressure, while the 6GB VRAM guide focuses on constrained cards.

6. Build a Minimal Reproduction

Replace the complex graph with a default checkpoint-to-sampler-to-VAE path. Add custom nodes back one branch at a time. Keep the same model and seed. If the minimal graph works, the GPU and core installation are probably healthy; the failing branch becomes your search area.

7. What to Include in a Bug Report

• Operating system, GPU, driver, ComfyUI version, Python and PyTorch versions.
• Installation method: Desktop, Portable, manual, or CLI.
• Full error text and a minimal workflow JSON.
• Custom-node list and the last change before failure.
• Expected result and exact reproduction steps.

8. Recovery Playbook by Symptom

Blank interface after an update

Test the core interface with third-party frontend extensions disabled. If it loads, restore extensions in small groups. Do not erase the installation before recording the versions that produced the failure.

Workflow opens with red missing nodes

List the exact missing node types, use Manager's missing-node view, and match each node to its original repository. Install only the packs required by the workflow. A node with a similar display name may accept different inputs and is not automatically a safe replacement.

Generation reaches the sampler and crashes

Reduce image dimensions and batch size, then repeat with the same seed. If the smaller job succeeds, investigate peak memory at that node. If it fails identically, examine model compatibility and the traceback rather than assuming the card is too small.

API call returns success but no image appears

Store the returned prompt ID, follow execution messages, and inspect the history/output route. A successful queue submission proves validation passed; it does not prove every downstream node completed.

9. Prevent the Next Failure

• Keep a known-good workflow that uses only core nodes.
• Export production graphs before every update.
• Record model filenames and checksums for critical projects.
• Update core and custom nodes in separate maintenance windows.
• Keep input data and generated deliverables outside the application folder.
• Review logs after recovery and write down the actual cause.

Frequently Asked Questions

Why does Queue Prompt do nothing?

Check validation messages and the backend log first. A missing model or invalid required input can prevent the workflow from entering the queue.

Should I update every custom node at once?

Not during diagnosis. Change one layer at a time so you know which change fixed or caused the issue.

Is clearing VRAM always the fix?

No. Missing nodes, invalid workflows, dependency conflicts, and broken WebSockets are not VRAM problems.

Primary references: ComfyUI official troubleshooting, custom-node management, server overview, and route documentation.