ThinkFacility

Error messages

Reconnecting

The message

Reconnecting
ComfyUI 1.55.11 read September 25, 2026ComfyUIout of memoryWebSocket

What it means

The page in your browser lost its live connection to the ComfyUI server and is trying to get it back. When it happens mid-generation, the server process has almost always stopped, and running out of memory is the cause ComfyUI's maintainers point to first.

What to do

Look at the terminal or console window ComfyUI runs in: the last lines before it closed or restarted tell you why. For memory, use a smaller or GGUF model, lower the resolution or length, and give Windows a bigger page file.

This is the red toast in the corner of the ComfyUI window, usually a few seconds into a run:

Reconnecting

Older builds spelled it Reconnecting..., and when the connection comes back it's replaced by a green Reconnected. We read the frontend code in version 1.55.11. The browser page keeps a WebSocket open to the ComfyUI server (a live two-way connection it uses for progress and previews). When that socket closes, the page tries a new one every 300 milliseconds, and if it hasn't got one back within two seconds, it shows this toast.

What usually closed the connection

The page can't tell you why the socket closed, only that it did. When it happens during a generation, it's almost always because the Python server on the other end stopped. The ComfyUI maintainers keep landing on the same answer: on a video workflow in issue #9205, ltdrdata wrote "It appears to have been forcibly terminated due to insufficient memory", and on a Z Image workflow in #10992, comfyui-wiki's first guess was "Maybe it's an OOM issue" (out of memory).

The memory that runs out can be ordinary system RAM as well as the graphics card's. Big models get loaded or moved through RAM, and when it's gone the operating system kills the process outright, without the neat Python error you'd get for a full GPU. That's why there's no error dialog, just a dropped connection. It surprised us how often the fix in those threads was on the Windows side: in #10992 one user said raising the page file (disk space Windows uses as overflow RAM) to 24 GB stopped it, though another said it didn't help them.

Finding out which it was

Go to the terminal window ComfyUI runs in (the black console for the portable build, the logs for the desktop app). A ComfyUI team member asked for exactly that in #9205. What you're looking for is the last thing printed before the window closed or the server restarted. A Python traceback there points to a node or custom node that crashed. If the output just stops mid-step with nothing, the process was killed from outside, and memory is the likely reason.

What helps

For memory, the advice in those threads is to shrink the job: comfyui-wiki suggested a GGUF version of the model (a compressed format that needs less memory) to a user short on VRAM, and lower resolution or fewer frames help on video models. Close other heavy programs, give Windows a larger page file on a drive with room, and restart ComfyUI fully if a previous run left Python processes behind.

If the toast keeps appearing and clearing while jobs carry on running, the server is fine and the connection itself is flaky. In #15016, users on Modal traced repeated Reconnecting and Reconnected toasts to a bug in a recent release of aiohttp, the web library ComfyUI's server uses. There's a setting that hides these toasts, "Disable toasts when reconnecting or reconnected", but the frontend marks it hidden, and it only hides the message.

If the page says Disconnected from backend. Check if the server is running. instead, the server never came back. Start it again from the terminal and read what it printed last time.

Other lines the same feature prints

Match yours against these if the one at the top of the page is not quite it. They come from the same code and mean related things.

  • Reconnecting...
  • Reconnected
  • Disconnected from backend. Check if the server is running.