ComfyUI usually breaks for one of four reasons: the server will not launch, missing custom nodes show up red, the browser is stuck in a Reconnecting websocket loop, or checkpoints do not appear in the loader. Almost every case is fixed by installing missing nodes through Manager, correcting a model folder path, or restarting the server cleanly.
ComfyUI is a node graph application, so its failures look different from a one click generator. When something is wrong you get red nodes, console tracebacks, and a canvas that will not run, rather than a simple error toast. That can feel intimidating, but it also means the cause is almost always readable in two places: the terminal window where the server runs, and the color of the nodes on your canvas. This guide covers every common 2026 breakage and the exact fix, in the order you should check them.
If you are still deciding how deep to go with a local node based workflow, the ComfyUI for NSFW AI complete guide covers setup and graph building from scratch, and the ComfyUI versus Forge versus A1111 comparison explains where each front end fits. This article assumes ComfyUI is installed and has stopped working.
First: read the terminal, not just the browser
The browser canvas only shows you symptoms. The terminal window (the black console that opened when you launched ComfyUI) shows you causes. Before you change anything, scroll the terminal to the last dozen lines and look for the words “Traceback”, “Error”, “failed”, or “import”. Ninety percent of ComfyUI problems announce themselves there in plain English. Keep that window visible while you troubleshoot; every fix below is confirmed or ruled out by what it prints.
| Symptom on screen | What the terminal shows | Root cause |
|---|---|---|
| Server never opens the browser page | Traceback ending in a module name | A dependency or custom node failed to import |
| Nodes are red with “missing” text | “cannot find node type” on load | Custom nodes referenced by the workflow are not installed |
| Canvas says “Reconnecting…” forever | Nothing new, or a repeating disconnect line | Server crashed or the websocket dropped |
| Checkpoint loader dropdown is empty | “0 models found” style lines | Models are in the wrong folder or path is wrong |
| Custom node fails with “IMPORT FAILED” | A red import traceback for that node pack | Missing python dependency for that pack |
The server will not launch
If ComfyUI never opens the page and the terminal closes or throws a traceback, the launch itself is failing. Work through these in order.
First, confirm you are launching the right way for your install type. A portable ComfyUI build is started with its bundled batch file such as run_nvidia_gpu.bat, which uses its own embedded python. A manual install is started from your activated virtual environment with a python main.py command. Mixing them, for example running main.py with your system python against a portable install, guarantees import errors because the dependencies live in different places. Always launch a portable build with its own bat file and a manual build from its own venv.
Second, read the traceback for the failing module. A line like “No module named” followed by a package name means a dependency is missing. In a manual install, activate the venv and reinstall requirements. In a portable install, use the embedded python to install into the right site packages rather than your system python. A traceback that names a specific custom node folder means that node pack, not ComfyUI core, is what is crashing the launch. Temporarily move that pack out of the custom_nodes folder and relaunch; if it starts, that pack is the culprit and needs updating or reinstalling.
Third, if the launch fails after a git pull or an update, a dependency version drifted. Update ComfyUI and its Manager together, then let Manager reinstall dependencies. A clean relaunch after an update fixes most post update launch crashes.
Red missing custom nodes and how to install them
Red nodes with a “missing” or “cannot find node type” label are the single most common ComfyUI problem, and they are usually trivial to fix. They appear when you load a workflow that was built with custom nodes you do not have installed. The workflow file references node types by name; if the node is not present, ComfyUI draws it red.
The fix is ComfyUI Manager. If you do not have Manager, install it first, because it is the tool that makes node management sane. With Manager present:
- Open Manager from the button on the ComfyUI menu.
- Choose “Install Missing Custom Nodes”. Manager scans the current workflow and lists every node pack it needs but cannot find.
- Install the listed packs, then fully restart the ComfyUI server, not just the browser tab.
- Reload the workflow. The previously red nodes should now resolve.
If a node stays red even after installing, the pack that provides it may have been renamed or deprecated, or two packs provide a conflicting node of the same name. Check the Manager list for a newer replacement pack, and remove duplicate packs that define the same node type. Restart the server after any node change; ComfyUI only re scans custom nodes at startup.
“IMPORT FAILED” custom nodes
A node pack that shows “IMPORT FAILED” in red in the Manager list, or throws an import traceback in the terminal at startup, installed its files but could not load because one of its own python dependencies is missing or incompatible. This is different from a missing node; the files are there, they just will not import.
Open Manager and use the “Try Fix” option on the failed pack, which reinstalls that pack’s requirements. If that does not clear it, read the terminal traceback for the exact missing module and install it into the correct python environment (the venv for a manual install, the embedded python for a portable one). Occasionally a pack pins a dependency version that clashes with ComfyUI core; in that case, update both ComfyUI and the pack to their current versions so their requirements line up again. If a single stubborn pack keeps breaking every launch and you do not need it, removing its folder from custom_nodes is a legitimate fix.
The “Reconnecting” websocket loop
When the canvas overlay says “Reconnecting…” and never clears, the browser has lost its websocket link to the ComfyUI server. There are two distinct causes and they need different fixes.
If the terminal window has closed or shows a fresh traceback, the server actually crashed, usually mid generation because it ran out of memory or a custom node threw an exception. The browser is faithfully trying to reconnect to a server that is no longer there. Relaunch the server; the fix is on the server side, and if it crashed on out of memory you also need the memory section below.
If the terminal is still running fine but the browser still says “Reconnecting”, the link between browser and a live server broke, not the server itself. This is a networking or proxy problem. Refresh the page. If you are reaching ComfyUI through a reverse proxy, a tunnel, or a remote address, confirm that proxy forwards websocket connections, because a proxy that only forwards HTTP will let the page load but never keep the socket alive. Browser extensions and strict privacy settings can also kill the websocket, so test in a clean incognito window.
| “Reconnecting” scenario | Tell tale sign | Fix |
|---|---|---|
| Server crashed | Terminal closed or new traceback | Relaunch server, address the crash cause |
| Out of memory crash | Terminal shows an allocation error | Relaunch, lower resolution or batch, tune VRAM |
| Proxy or tunnel dropping socket | Terminal healthy, remote access | Ensure the proxy forwards websockets |
| Extension blocking socket | Works in incognito | Whitelist the address or disable the extension |
App level out of memory
ComfyUI crashing or refusing to finish a render with a memory error is a resource problem, not a bug. At the application level you have several levers before touching anything deeper. Lower the output resolution, since VRAM cost scales with pixel count and a large image can exhaust an otherwise capable card. Reduce the batch size to one. Add the low VRAM or normal VRAM launch flag appropriate to your card so ComfyUI manages memory more conservatively. Enable tiled VAE decoding for large images, which decodes in chunks instead of all at once and avoids the final step memory spike that causes many out of memory crashes. If your goal is large final images, generate at a safe base size and enlarge afterward using the workflow in how to upscale NSFW images in ComfyUI, which is far less likely to crash than rendering huge in one pass.
For the deeper, output quality side of VRAM tuning, including the exact settings that let you push resolution without a crash, the CUDA out of memory fix is the companion piece and goes further than the app level flags here. If your card is simply too small for the resolutions you want, that guide also covers the realistic ceilings per VRAM tier.
Checkpoints and models not showing in the loader
An empty checkpoint loader dropdown, or a LoRA or VAE that will not appear in its node, is nearly always a folder path problem, not a corrupted model. ComfyUI only lists models it finds in the folders it is told to scan.
Confirm your checkpoint files sit in the ComfyUI models checkpoints folder, LoRAs in models loras, and VAEs in models vae. Files dropped loosely into the ComfyUI root or into a subfolder ComfyUI is not scanning will not show. After adding a model, restart the server, because ComfyUI reads the model folders at startup and does not live watch them; a new file added while the server runs will not appear until you either restart or use the refresh option. If you keep a shared model library used by another front end, ComfyUI supports an extra model paths configuration file that points it at those external folders, which avoids duplicating large checkpoint files. Set that up correctly and your existing library appears without copying anything. If a specific safetensors file shows but fails to load, it may be corrupt or a partial download, so re download it and compare the file size. If you are unsure which models are worth keeping in that folder, the roundup of best Stable Diffusion checkpoints for NSFW covers the current reliable options.
Portable versus manual install: which problems belong to which
Many ComfyUI headaches come down to confusing the two install styles. A portable build ships a self contained embedded python and is the more forgiving option for most users, because its dependencies are isolated and its launch bat files are preconfigured. A manual install uses a python virtual environment you created and gives you more control at the cost of managing that environment yourself.
The practical rule: always install packages into the same python that runs ComfyUI. For a portable build that means the embedded python, not your system python. For a manual build that means the activated venv. Nearly every “No module named” error on an otherwise complete install traces back to a package installed into the wrong python. When in doubt, a portable build sidesteps most of these environment problems and is the recommended starting point if you are fighting dependency errors repeatedly.
Updating safely without breaking a working install
A large share of ComfyUI breakages happen right after an update, so update deliberately rather than impulsively. Update ComfyUI core and ComfyUI Manager together, then let Manager reconcile node pack dependencies, because a core update paired with stale node packs is the classic recipe for a launch crash. Before updating custom nodes in bulk, note which packs you actually rely on; updating a pack you never use only adds risk. If an update does break the launch, the terminal traceback will name the offending pack or module, and you can roll that single pack back or move it aside rather than reinstalling everything. Portable builds make this safer still, since you can keep the previous portable folder untouched and test the update in a copy before committing to it. Treating updates as reversible experiments, not one way commitments, is the habit that keeps a working ComfyUI working.
If ComfyUI keeps breaking, a hosted option skips the whole problem
Local ComfyUI is powerful, but it is also a moving target: node packs update, dependencies drift, and a single bad git pull can cost an evening. If you find yourself spending more time fixing the install than generating, a hosted no install NSFW generator sidesteps the entire setup, since there are no nodes, no VRAM limits, and no environment to maintain. A service like AI Nudez runs everything server side, so you get straight to generating while your local ComfyUI waits for a calmer day. Many people keep both: the hosted tool for quick results and ComfyUI for full control when the setup is behaving.
When to reinstall clean
Reinstalling is a last resort, but sometimes an install accumulates so many conflicting node packs and pinned dependencies that a clean slate is faster than untangling it. Before you do, back up your models folder and your important workflow JSON files, because those are the parts worth keeping. Then set up a fresh portable ComfyUI, reinstall only the node packs you actually use through Manager, and copy your models back or point the extra model paths file at your existing library. Add packs one at a time and launch after each so that if something breaks you know exactly which pack did it. That disciplined rebuild almost always ends the cycle of launch crashes.
For a broader diagnostic framework that applies across every local and hosted tool, keep the NSFW AI troubleshooting guide handy; it maps the same read the console, isolate the cause, apply the matching fix logic to the whole ecosystem rather than to ComfyUI alone.
Frequently asked questions
Why won’t ComfyUI launch or open the browser page?
The launch is failing, and the terminal traceback names the cause. A missing module means a dependency did not install, so reinstall requirements into the correct python. A named custom node folder in the traceback means that pack is crashing startup, so move it out of custom_nodes and relaunch. If it broke right after a git pull, update ComfyUI and Manager together to realign dependency versions.
How do I fix red missing custom nodes in ComfyUI?
Red nodes mean the workflow references custom nodes you have not installed. Open ComfyUI Manager, choose Install Missing Custom Nodes, install everything it lists, then fully restart the server, not just the browser tab. Reload the workflow and the red nodes resolve. If a node stays red, the pack was renamed or two packs conflict, so find the current replacement and remove duplicates.
Why does ComfyUI say Reconnecting and never recover?
The browser lost its websocket to the server. Check the terminal: if it closed or shows a new traceback, the server crashed, often on out of memory, so relaunch and address the crash. If the terminal is still healthy, the socket dropped between browser and server, which points to a proxy that is not forwarding websockets or a blocking extension. Refresh, or test in incognito.
Why are my checkpoints not showing in the ComfyUI loader?
ComfyUI only lists models in the folders it scans. Put checkpoints in models checkpoints, LoRAs in models loras, and VAEs in models vae, then restart the server, because it reads those folders only at startup. To reuse an existing library from another front end, set up the extra model paths configuration file so ComfyUI scans your external folders without copying the files.
What does IMPORT FAILED mean on a ComfyUI custom node?
The node pack installed its files but could not load because one of its own python dependencies is missing or incompatible. In ComfyUI Manager, use Try Fix on that pack to reinstall its requirements. If it persists, read the terminal for the exact missing module and install it into the python that runs ComfyUI. Updating both ComfyUI and the pack often realigns clashing dependency versions.
How do I stop ComfyUI running out of memory?
Lower the output resolution, since VRAM cost scales with pixel count, and set batch size to one. Add the low VRAM launch flag for your card, and enable tiled VAE decoding so the final decode happens in chunks rather than one memory spike. For deeper output quality VRAM tuning that pushes resolution without crashing, follow the dedicated CUDA out of memory fix guide, which goes beyond the app level flags.
What is the difference between portable and manual ComfyUI for fixing errors?
A portable build ships its own embedded python with isolated dependencies and preconfigured launch bat files, so it is more forgiving. A manual install uses a virtual environment you manage. The key rule when fixing errors is to install packages into the same python that runs ComfyUI: the embedded python for portable, the activated venv for manual. Most No module named errors come from installing into the wrong python.
Should I reinstall ComfyUI when it keeps breaking?
Reinstalling is a last resort but sometimes faster than untangling conflicting node packs. First back up your models folder and important workflow JSON files. Set up a fresh portable build, reinstall only the packs you use through Manager, and add them one at a time, launching after each so you can spot which pack breaks. Point the extra model paths file at your existing library instead of copying checkpoints.
