When ComfyUI keeps reconnecting, start with the server process. The page can remain open after the process has stopped, crashed or become unreachable. Repeatedly reloading the browser does not explain which of those happened.
Record when the connection disappears
Write down the last action before the message: starting the application, loading a particular graph, beginning model loading, sampling, or saving a large video. The timing narrows the investigation. A failure before any graph runs points toward a different area than a failure after increasing video resolution.
| When it happens | First check |
|---|---|
| Immediately after launch | Server startup output, selected instance and local address |
| After closing a terminal | Whether the terminal was running the server |
| During a heavy generation | Memory errors and the last completed node |
| After an extension update | The extension’s import or frontend errors |
| Only through a remote connection | The remote host and connection path |
Check the server before the browser
Look at the terminal or application log used to launch ComfyUI. Is the process still running? Does it show a traceback, an out-of-memory message, or a normal listening address? For Portable, the official instructions explain that its command window must remain open while the server runs.
If the process stopped, save the error text before restarting. Then use your normal launcher once. If startup fails again, the new log should tell you more than the reconnecting banner. Avoid starting several competing instances while trying to identify the active one.
Run a small known working graph
Use the baseline saved when you first installed ComfyUI. Keep batch size small and retain that example’s documented settings. If the baseline works but a larger video graph fails, compare the requirements and the logged error. Do not conclude that the entire installation is broken.
For memory-related failures, reduce the test workload and follow the model’s documented memory options. Lowering output size or batch size is a diagnostic change, not proof that every model can run on your hardware. Keep the failed and successful settings in your notes.
Isolate extension problems
ComfyUI’s official troubleshooting process includes starting with custom nodes disabled and re-enabling them in groups to find the conflicting package. If a core-only graph succeeds while the extended setup fails, investigate that difference before replacing model files.
A useful comparison uses the same installation and the same small graph. Change only whether the suspected extension is active. An experiment that changes the browser, model, graph and Python version together cannot identify which change mattered.
What “failed to fetch server logs” tells you
It tells you the interface could not retrieve those logs through that path. It does not establish the original cause. Read the local launch terminal or the relevant application log directly, and verify that you are looking at the same instance the page is trying to reach.
For a remotely hosted setup, record the host status and whether the issue also occurs when using its intended access path. Keep authentication and network protections in place. Broadly disabling protections or exposing the server publicly is not a diagnosis.
“Prompt has no outputs” is a different error
When the server responds but a graph has no usable output, inspect the graph’s save or preview endpoint and its connected inputs. Start from an intact official template and compare the connections. Do not treat a validation message as proof of a connection failure. The workflow guide explains how to trace the graph from inputs to outputs.
Prepare a useful issue report
Include the installation method, software version, operating system, GPU, smallest reproducing graph, first relevant error and the last change before failure. Remove private prompts and reference files unless they are necessary and appropriate to share. This gives a maintainer a specific problem to reproduce.
Once the baseline works, export and back up the workflow. If your immediate goal is creator content rather than local setup maintenance, create a persona in Clout and evaluate a hosted image and video workflow.
