You installed ComfyUI, double-clicked it, and instead of images you got a red message, or the black window closed by itself. It happens to almost everyone the first time, and it still happens to us now and then after an update: we work with ComfyUI every day on a perfectly ordinary 12 GB RTX 3060, and we have seen plenty of these errors.
The good news is that the cause is almost always simple: a missing Windows library, an old driver, a file in the wrong folder. In this guide we have collected the most common problems people run into when they install ComfyUI for the first time, both with the portable version and with the Desktop app, with step-by-step fixes. We checked everything against the official documentation and the bug reports of the last few months (the guide is up to date as of November 2026, ComfyUI 0.46).
How to use this page: copy the error message you see, press Ctrl+F and search for part of it below. The errors are written exactly as ComfyUI shows them.
At a glance: the error and where to find the fix
| If you see… | It usually means | Go to |
|---|---|---|
c10.dll, WinError 126, MSVCP140.dll | The Microsoft Visual C++ libraries are missing | Microsoft libraries |
Torch not compiled with CUDA enabled | Old NVIDIA driver or wrong package | NVIDIA driver |
no kernel image is available | The package doesn't fit your graphics card | The right package |
| The black window closes right away | A startup error you don't have time to read | Window that closes |
Port 8188 is already in use | ComfyUI is already open | Port in use |
Red nodes, node types were not found | Some custom nodes are missing | Red nodes |
Value not in list: ckpt_name | The model is missing or in the wrong folder | Missing models |
CUDA out of memory | Your graphics card doesn't have enough memory | Video memory |
| Black or gray image | Wrong VAE or a precision problem | Black images |
mat1 and mat2 shapes cannot be multiplied | Models from different families mixed together | Incompatible models |
| The Desktop app opens and closes | An app update stopped halfway | App that closes |
First things first: where to read the error
To fix a problem we need the exact message. Where to find it depends on how you installed ComfyUI.
Portable version. The error shows up in the black window that opens with run_nvidia_gpu.bat. If there is a long list of lines starting with Traceback, the line that matters is almost always the last one, or the first one with the word Error.
Desktop app. The app's messages are in the menu (the three lines at the top) → Desktop Settings → Logs → Open logs folder (the file is called app.log). The messages from ComfyUI itself are in your installation's card: the three dots → Manage → Terminal tab.
Error in the interface. If ComfyUI starts but a node turns red, the message appears in a box in the browser and also in the black window (or in the app's Terminal).
Portable version problems
The portable version is the .7z package you download from GitHub, extract and start with run_nvidia_gpu.bat. It's the one most guides use, and it's also the one that causes the most trouble the first time, because it relies on the libraries and drivers already on your PC.
"c10.dll", "WinError 126" or "MSVCP140.dll not found" error: the Microsoft libraries are missing
Typical messages:
OSError: [WinError 126] The specified module could not be found. Error loading "...\torch\lib\c10.dll" or one of its dependencies.DLL load failed, or a Windows window saying thatMSVCP140.dllorVCRUNTIME140.dllis missing
Why it happens: PyTorch, the engine that runs the models, needs a package of Microsoft libraries called Visual C++ Redistributable. Many PCs already have it because other programs or games installed it, but on a new or freshly formatted PC it's often missing.
How to fix it:
- Download the official Microsoft package from this address: https://aka.ms/vc14/vc_redist.x64.exe. It's the same link ComfyUI writes in its startup file.
- Open the downloaded file and click Install. If it says it's already installed, choose Repair: sometimes the existing installation is damaged.
- Restart the computer.
- Run
run_nvidia_gpu.batagain.
You will still find the old link aka.ms/vs/17/release/vc_redist.x64.exe around: it works, but today Microsoft points to the one above. If the error is still there after the installation, the most likely cause is the graphics card driver (see the next section).
"Torch not compiled with CUDA enabled" or "The NVIDIA driver on your system is too old"
Why it happens: The standard portable package uses a version of PyTorch built for CUDA 13, which needs a recent NVIDIA driver (the 580 series or later). With an older driver ComfyUI can't see the graphics card. The same error shows up if you have an AMD or Intel card and downloaded the NVIDIA package.
How to fix it:
- Check the driver version: press Windows + R, type
cmd, press Enter and in the window typenvidia-smi. The driver number is on the first line at the top (for example "Driver Version: 581.xx"). - If the number is lower than 580, install the latest driver from the NVIDIA app or from nvidia.com/drivers. Both Game Ready and Studio drivers are fine.
- Restart the PC and run
run_nvidia_gpu.batagain. - If you have an AMD or Intel Arc card, the NVIDIA driver has nothing to do with it: download the right package for your card (see below).
NVIDIA doesn't publish an official minimum version for Windows: the "580" comes from the driver series tied to CUDA 13. When in doubt, just install the latest driver available.
"no kernel image is available" or "sm_61 is not compatible": the package doesn't fit your card
Typical messages: NVIDIA GeForce GTX 1060 with CUDA capability sm_61 is not compatible with the current PyTorch installation followed by CUDA error: no kernel image is available for execution on the device. With RTX 50 series cards the number becomes sm_120.
Why it happens: The ComfyUI download page offers different packages, and each one only works with certain cards. The standard package no longer supports GTX 10 series cards and older ones; the package for older cards, on the other hand, doesn't support the RTX 50 series.
Which package to download (from the ComfyUI releases page on GitHub):
| Your graphics card | File to download |
|---|---|
| NVIDIA RTX 20, 30, 40, 50 and GTX 16 | ComfyUI_windows_portable_nvidia.7z |
| NVIDIA GTX 10 (1060, 1070, 1080…) and older | ComfyUI_windows_portable_nvidia_cu126.7z |
| AMD Radeon RX 6000, 7000, 9000 | ComfyUI_windows_portable_amd.7z (starts with run_amd_gpu.bat) |
| Intel Arc | ComfyUI_windows_portable_intel.7z (starts with run_intel_gpu.bat) |
How to fix it:
- Check which card you have: press Ctrl + Shift + Esc to open Task Manager, go to Performance and click GPU. The card's name is at the top right.
- Download the right package from the table.
- Extract it into a new folder (see the tips below) and run the matching
.batfile. - With RTX 50 series cards, don't reuse an old portable package downloaded before 2025: it doesn't include support for these cards.
Extracting the .7z file doesn't work or some folders are missing
Symptoms: 7-Zip reports Data error or Cannot open the file as archive, or when you start ComfyUI you get '.\python_embeded\python.exe' is not recognized.
Why it happens: The file is about 2 GB, so if the download was interrupted it's incomplete. Or Windows blocked the file because it comes from the internet.
How to fix it:
- Compare the size of the downloaded file with the one shown on GitHub. If it's different, download it again.
- Install 7-Zip, which is free: older versions of File Explorer don't open
.7zfiles properly. - If extraction still fails: right-click the
.7zfile → Properties → at the bottom check Unblock → OK. - Extract with right-click → 7-Zip → Extract Here, into a folder with a short path.
- Check that it contains the
ComfyUI,python_embededandupdatefolders and therun_nvidia_gpu.batfile.
Where to put the ComfyUI folder
Many strange problems (files that disappear, permission errors, nodes that won't install) come from where the folder is. Our advice:
- use a short path on a local drive, for example
C:\AI\ComfyUI_windows_portableorD:\AI\...; - avoid the Desktop, Documents and Downloads if OneDrive is active on your PC, because those folders get synced to the cloud;
- if you get the
Filename too longerror while installing nodes, Windows has a 260-character limit for paths: moving the folder higher up (for example straight intoC:\AI) is usually enough.
For more experienced users, there is also a Windows setting that removes the long path limit (Group Policy → "Enable Win32 long paths"), but it needs administrator access, and for a beginner moving the folder is simpler.
The black window opens and closes right away
Why it happens: There's an error at startup, but the window closes before you can read it. The original run_nvidia_gpu.bat file ends with a command that keeps the window open: if it closes anyway, you've usually copied the .bat file to another folder (for example the Desktop) or edited it.
How to fix it:
- Open the ComfyUI portable folder.
- Click the address bar at the top of File Explorer, type
cmdand press Enter: a window opens already in the right folder. - Type
run_nvidia_gpu.batand press Enter. Now the window stays open and you can read the error. - If you want to start ComfyUI from the Desktop, don't copy the
.batfile: create a shortcut (right-click the file → Show more options → Create shortcut) and move that instead. The.batfile only works from its own folder.
"Port 8188 is already in use"
The full message is Port 8188 is already in use on address 127.0.0.1. Please close the other application or use a different port with --port.
Why it happens: ComfyUI is already open, for example because you double-clicked twice, or the Desktop app is open too.
How to fix it:
- Close all ComfyUI black windows and the Desktop app, if it's running.
- If the error is still there, open Task Manager (Ctrl + Shift + Esc), look for any
python.exeprocesses left open and close them with End task. - If you want to run two ComfyUI instances at once, make a copy of the
.batfile, open it with right-click → Edit and add--port 8189at the end of the first line. Then openhttp://127.0.0.1:8189in the browser.
The browser doesn't open or the page stays blank
How to fix it:
- Wait until the line
To see the GUI go to: http://127.0.0.1:8188appears in the black window. - Open Chrome or Edge and type
http://127.0.0.1:8188by hand. - Update your browser: the official documentation lists Chrome 143 as the minimum, and with older versions the interface may not load.
- If you use a VPN or a company proxy, try turning them off.
The Manager isn't there: I can't install nodes
Why it happens: Since December 2025 the ComfyUI Manager (the tool for installing nodes and models) is built into ComfyUI, but in the portable version it's turned off: you have to install and enable it once. Many older guides tell you to download it with git clone into the custom_nodes folder: that still works, but today this is the official method.
How to fix it:
- In the main ComfyUI portable folder, create a text file and name it
install_manager.bat(make sure the extension is.batand not.txt). - Open it with Notepad and write two lines:
.\python_embeded\python.exe -m pip install -r ComfyUI\manager_requirements.txton the first,pauseon the second. Then save. - Double-click
install_manager.batand wait for it to finish. - Make a copy of
run_nvidia_gpu.batand name it, for example,run_with_manager.bat. - Right-click the copy → Edit, and change the first line to this:
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-manager - From now on, start ComfyUI with
run_with_manager.bat. In the interface you'll find the Manager under the puzzle icon or in the menu → Manage Extensions.
In the Desktop app the Manager is already active, so there's nothing to do.
The antivirus deletes some files
If some files are missing after extraction or Windows shows "Threat quarantined", open Windows Security → Virus & threat protection → Protection history and see which file was blocked. Restore it only if it's part of the official package downloaded from github.com/Comfy-Org. Don't turn off the antivirus for the whole folder: custom nodes are third-party code, and in the past some of them really did contain malware.
Desktop app problems
Watch out for old guides: in 2026 the official app was rebuilt from scratch and is now called Comfy Desktop (version 1.x). It has a start screen with your installations, called "instances", and automatic backups. The maintenance page, the "Reinstall" button and the Git and Visual C++ checks you see in many tutorials belong to the old app, which is now called Legacy Desktop.
You can download the app from comfy.org/download. On Windows it works with NVIDIA cards, and there are also versions for AMD (RX 6000 and later), Intel Arc and CPU only. On the Mac it needs an Apple Silicon chip (M1 or later) with macOS 13 or newer: Macs with Intel processors are not supported.
The app opens for a second and closes (after an update)
Why it happens: The app's automatic update stopped halfway. It's a known problem in versions up to 1.0.46.
How to fix it:
- Download the latest installer from comfy.org/download.
- Run it over the existing app, without uninstalling it first: installations, models and settings stay where they are.
- If that's not enough, here is a fix reported by several users, though not an official one: close the app, open the file
%APPDATA%\Comfy Desktop\settings.jsonwith Notepad (paste the path into the File Explorer address bar) and change"installUpdatesOnStartup": truetofalse. Save and reopen the app.
The app doesn't open at all
- On Windows: right-click the icon → Run as administrator, only the first time.
- On the Mac: System Settings → Privacy & Security → at the bottom click Open Anyway.
- Make sure you have the latest version of the app.
"This file was blocked by application control policy" (Windows 11)
Typical messages: ImportError: DLL load failed while importing ...: This file was blocked by application control policy.
Why it happens: On some Windows 11 PCs Smart App Control is active, a protection that blocks files without a digital signature, and some Python libraries don't have one.
What to know: Right now (November 2026) there is no official fix: the problem has been reported to the developers. The only way to get it working is to turn off Smart App Control in Windows Security → App & browser control → Smart App Control settings. It's a security protection, and once it's off Windows only lets you turn it back on by reinstalling the system: decide for yourself whether it's worth it, or use the portable version.
The first download gets stuck or "ERR_CONNECTION_RESET" appears
Why it happens: On the first installation the app downloads several GB (the Python environment and PyTorch). If the connection drops, the download fails.
How to fix it:
- Use a stable connection, preferably wired; turn off VPNs and phone hotspots.
- Try creating the installation again: the app reuses the part already downloaded.
- In the menu (the three lines at the top) → Desktop Settings → Advanced, check that PyPI Mirror URL is empty and Use Chinese Mirrors is off (they're only needed from China).
"Operation failed" screen
- Press Reboot (or Restart ComfyUI) to try again.
- If it still fails, open the logs (the menu (the three lines at the top) → Desktop Settings → Logs) and the Terminal tab in Manage to read the real error, then search for it on this page.
- If the installation stays broken: the three dots on its card → Uninstall, then create a new installation with New Instance. The models are in a shared folder and don't get deleted.
"Torch not compiled with CUDA enabled" or "access violation" when the app starts
- Update the app to the latest version.
- Update the NVIDIA driver: the app itself shows a warning if the driver is older than the 580 series.
- If the error is still there: the three dots → Manage → Update tab → in the PyTorch section choose another version and press Change PyTorch. With Copy & Change PyTorch you can test it on a copy, without touching the original installation.
I have a GTX 10 series card
With the GTX 1060, 1070, 1080 and similar cards you get the warning "needs compute capability 7.5" or, when generating, no kernel image is available. In Manage → Update → PyTorch choose the cu126 version (the one for older cards) and press Change PyTorch.
I can't choose the installation folder
If the Continue button stays gray, the app is rejecting the folder you picked: it doesn't accept OneDrive folders or folders already used by another installation. Pick a normal folder on a local drive, for example D:\ComfyUI-Installs. You can change the default folder in the menu (the three lines at the top) → Desktop Settings → Advanced → Default Install Location.
Where my files are in the Desktop app
| What | Windows | Mac |
|---|---|---|
| Installations | %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Installs | ~/ComfyUI-Installs |
| Shared models, input and output | %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared | ~/ComfyUI-Shared |
| Settings and logs | %APPDATA%\Comfy Desktop | ~/Library/Application Support/Comfy Desktop |
To open these folders on Windows, paste the path into the File Explorer address bar. If you already had a portable version, you can add it to the app with the menu (the three lines at the top) → Add Existing Instance, without copying anything.
The first workflow: red nodes and missing models
ComfyUI starts, you load a workflow downloaded from the internet and something turns red. These are by far the most frequent problems, and the easiest to fix.
"When loading the graph, the following node types were not found"
Why it happens: The workflow uses custom nodes (extensions made by the community) that aren't installed on your ComfyUI.
How to fix it:
- Make sure the Manager is active (in the portable version you have to enable it, see above).
- When you load the workflow, ComfyUI offers to install the missing nodes: choose Install All, or Open Manager to see them one by one.
- When it's done, close and restart ComfyUI completely (reloading the page isn't enough), then load the workflow again.
- If some nodes stay red, check the black window for
IMPORT FAILED(see the next section).
"IMPORT FAILED": the node is installed but won't load
Why it happens: The node is missing a Python library, or it conflicts with another extension or with your version of ComfyUI.
How to fix it:
- In the black window, scroll up above the
IMPORT FAILEDline: the first error you find (usuallyModuleNotFoundError: No module named ...) is the real cause. - Update that node in the Manager. If the Try Fix button appears, use it: it reinstalls the node's libraries.
- To find out whether the problem is a custom node, start ComfyUI without extensions: in the portable version make a copy of the
.batand add--disable-all-custom-nodesto the first line. If it works this way, one of the extensions is the culprit. - If a node keeps causing trouble, disable it from the Manager and look for an alternative.
"This action is not allowed with this security level configuration"
Why it happens: The Manager has a security level that blocks some installations it considers risky, for example from a GitHub address or from outside the official catalog.
How to fix it:
- In the portable version, open the file
ComfyUI\user\__manager\config.iniwith Notepad (in older versions it's inComfyUI\user\default\ComfyUI-Manager\config.ini; the exact path is written in the black window at startup). - Change the line
security_level = normaltosecurity_level = weak, save and restart ComfyUI. - Install what you need, then **set it back to
normal**: theweaklevel allows everything, including potentially dangerous installations. - In the Desktop app you do the same from Manage → Startup Args → Manager Security Level.
"Value not in list: ckpt_name": the model is missing
Typical message: Prompt outputs failed validation: CheckpointLoaderSimple: Value not in list: ckpt_name: 'model_name.safetensors' not in []
Why it happens: ComfyUI doesn't come with models: you have to download them separately. Or the model is there, but in the wrong folder, or the workflow is looking for a file with a different name from yours.
Which folder each model goes in (inside ComfyUI\models):
| File type | Folder |
|---|---|
| Full checkpoints (SD 1.5, SDXL…) | checkpoints |
| Separate diffusion models (Flux, Wan, Qwen-Image…) and GGUF files | diffusion_models (the old unet works too) |
| Text encoders (CLIP, T5, Qwen…) | text_encoders (the old clip works too) |
| VAE | vae |
| LoRA | loras |
| Upscaler | upscale_models |
| ControlNet | controlnet |
How to fix it:
- Put the file in the right folder from the table.
- In the interface press the R key to refresh the model list, or restart.
- Click the red node and pick the file from the drop-down menu: downloaded workflows carry the author's file name, which may be different from yours.
- In the Desktop app the models go in the shared folder (
ComfyUI-Shared\models), not in the program folder.
When you open a model from the ComfyUI Templates, the list of missing models appears with a button to download them. Sometimes it doesn't work: it's a known problem, reported to the developers and not fixed yet.
How to fix it:
- In the Desktop app, try again from the Downloads panel at the top.
- Otherwise download the file by hand from its Hugging Face page and put it in the right folder from the table above.
- On Hugging Face use the download button on the file's page: the right address contains
/resolve/main/. If the address contains/blob/main/, you're downloading the web page and not the model.
A workflow downloaded from the internet won't open
The workflow is saved inside the original PNG images and in .json files. If you drag in an image and only a "Load Image" node appears, that image doesn't contain the workflow: many websites and social networks convert images to JPG or WebP and strip the data. Look for the .json file or the original PNG, and open it with the File → Open menu instead of dragging it. Once it's open, nodes and models may be missing: see the sections above.
Memory and speed
"CUDA out of memory" or "Allocation on device"
Why it happens: The model, together with the resolution, the number of images or the length of the video, needs more memory than your graphics card has.
How to fix it, in this order:
- Lower the resolution, the number of images per run (batch size) or the length of the video.
- Close other programs that use the graphics card: browsers with lots of tabs, games, video editors.
- Use a lighter version of the model: fp8 or GGUF (compressed files that take up less memory, loaded with the ComfyUI-GGUF extension).
- Leave some memory free for Windows by adding
--reserve-vram 2to the first line of the.batfile (in the Desktop app: Manage → Startup Args).
Since March 2026 ComfyUI manages video memory automatically (Dynamic VRAM) on NVIDIA cards. That's why the old advice "add --lowvram" no longer does anything on its own today: it only works together with --disable-dynamic-vram. Try the steps above first.
On a 12 GB card like our RTX 3060, with fp8 or GGUF models you can easily make images with Flux and short videos: if you want to know how much memory you need for what, you'll find the details in our guide to graphics cards for AI. For video, our guide to AI video models in ComfyUI lists how much memory each model needs.
"The paging file is too small" (WinError 1455)
Why it happens: Here the graphics card has nothing to do with it: the computer's RAM has run out. Windows uses part of the disk as extra memory (the paging file), and the size it's set to is too small to load a large model.
How to fix it:
- Press Windows + R, type
sysdm.cpland press Enter. - Advanced tab → in the Performance section click Settings → Advanced tab → Virtual memory → Change.
- Uncheck "Automatically manage", select the fastest drive (ideally an SSD with free space), choose Custom size and set, for example, 32768 MB as both the initial and the maximum size.
- Click Set, OK and restart the PC.
Making the paging file bigger won't fix a CUDA out of memory error, though, because that one is about the graphics card.
If the error keeps coming back, your computer has too little RAM for the models you use. We work with 32 GB, and on our PC builds 2026 page you will find the computers we recommend for AI.
ComfyUI is extremely slow: it's using the processor
If an image takes many minutes and in Task Manager the graphics card stays idle, ComfyUI is working with the processor instead of the card. At startup the black window shows Device: cpu.
- Start ComfyUI with
run_nvidia_gpu.bat, not withrun_cpu.bat(since version 0.45 it has been moved to theadvancedfolder). - Check whether PyTorch sees the card: open
cmdin the ComfyUI portable folder and type.\python_embeded\python.exe -c "import torch; print(torch.cuda.is_available())". If it answersFalse, update the driver (see above) or download the right package again.
Sage Attention, Triton, xformers: you don't need them
Lots of videos promise faster generations with Sage Attention, Triton or xformers. They're optional optimizations, tricky to install on Windows, and they're among the most common reasons a ComfyUI install stops starting (for example with No module named 'sageattention'). If you've just installed ComfyUI, leave them alone. If you added them and something is wrong, remove --use-sage-attention from the .bat file and ComfyUI goes back to working with the standard method.
Wrong images and damaged files
Black or gray images
Why it happens: An all-black image, with no errors, usually comes from a calculation precision problem in the VAE (the part that turns the result into an actual image). A gray or brown image often means the VAE isn't the right one for the model.
How to fix it:
- Use the VAE that goes with the model: for SDXL the recommended file is
sdxl_vae_fp16_fix.safetensors, for Fluxae.safetensors. - Try adding
--fp32-vaeto the first line of the.batfile: it makes the VAE work with more precision. Careful: some guides mention--force-fp32-vae, which doesn't exist. - If it still doesn't work, start from an official ComfyUI template for that model: it has the right combination of files.
"mat1 and mat2 shapes cannot be multiplied"
Similar errors: expected input ... to have 4 channels, but got 16 channels instead, The size of tensor a must match the size of tensor b.
Why it happens: You're using files from different families together: for example an SDXL VAE with a Flux model, an SD 1.5 LoRA with an SDXL model, or the wrong text encoder.
How to fix it:
- Use the model, VAE, text encoder, LoRA and ControlNet from the same family (SD 1.5, SDXL, Flux, Qwen-Image…).
- If the error mentions 4 and 16 channels: SD 1.5 and SDXL work with 4 channels, Flux and SD3 with 16. You've mixed the two families.
- The safest way: open the model's official template from the ComfyUI Templates menu and replace only what you need.
To find out which files each model needs, you'll find the complete list in our guide to AI image models in ComfyUI.
"Error while deserializing header" (HeaderTooLarge)
Why it happens: The model file is damaged or incomplete. Sometimes it's actually a web page saved with a .safetensors name. This happens when you download the page address from Hugging Face instead of the file address, or from a page that requires you to log in.
How to fix it:
- Look at the file size: if it's a few KB instead of several GB, it's a web page.
- Download it again with the download button (address with
/resolve/main/). For models that ask you to accept a license, log in to Hugging Face and accept it first. - Always prefer
.safetensorsfiles over.ckptand.pt, which can contain executable code.
Something stopped working after an update
Which file to use to update the portable version
In the portable version's update folder there are three files:
| File | What it does | When to use it |
|---|---|---|
update_comfyui_stable.bat | Updates to the latest stable version | Our advice for beginners |
update_comfyui.bat | Updates to the latest development version | If you need a feature that just came out |
update_comfyui_and_python_dependencies.bat | Also updates PyTorch and all the libraries | Only if library errors show up: it's a big download |
Before a major update, make a copy of the ComfyUI folder (or at least of ComfyUI\custom_nodes and ComfyUI\user): if something breaks, you can go back in a moment.
NumPy errors after an update
Typical messages: A module that was compiled using NumPy 1.x cannot be run in NumPy 2.x or numpy.core.multiarray failed to import.
Why it happens: An update or a new node changed the version of NumPy (a math library) and an older extension doesn't support it.
How to fix it:
- Update the node that gives the error from the Manager: often the author has already fixed it.
- If that's not enough, disable that node or go back to the copy from before the update.
- Avoid changing the NumPy version by hand unless the node's author tells you to: you risk breaking other extensions.
Broken or outdated interface after the update
Symptoms: a white or gray page, menus out of place, or the warning Installed comfyui-frontend-package version ... is lower than the recommended version in the black window.
- Reload the page clearing the browser cache: Ctrl + F5.
- In the portable version, run
update\update_comfyui.batagain: it updates the interface too. - If the problem is still there, in Settings → Extensions disable the third-party extensions and turn them back on one at a time to find the one causing trouble.
- If your portable version is many months or years old, updating isn't enough: download a new package and move into it the
ComfyUI\models,ComfyUI\outputandComfyUI\userfolders and the nodes you need.
Going back to the previous version
- Desktop app: Manage → Snapshots. The app automatically saves a copy before and after every update (models are not included): restore the previous one. Next time use Copy & Update to try the update on a copy.
- Portable: if you made a copy of the folder, put it back in place. Otherwise, the Manager saves "snapshots" of the installed nodes that you can restore.
If nothing works: start over without losing your models
Sometimes the quickest route is a clean installation, and the models (which are the heavy part) don't need to be downloaded again.
- Portable: download a new package, extract it into a different folder and move
ComfyUI\models,ComfyUI\outputandComfyUI\user(that's where workflows and settings are) into the new folder. Reinstall the custom nodes from the Manager instead of copying them. - Desktop app: create a new installation with New Instance: the models are in the shared folder and the new installation sees them right away.
- Models in a single folder for several installations: in the portable version copy the file
extra_model_paths.yaml.example, rename itextra_model_paths.yamland write the path of your models folder inside it. In the Desktop app you do it from the menu (the three lines at the top) → Desktop Settings → Storage → Add Shared Directory.
If you're still deciding how to install ComfyUI, our guide to installing ComfyUI on Windows explains both methods from scratch, and the page ComfyUI: all our tutorials has everything else.
Frequently asked questions
Is the portable version or the Desktop app better for avoiding problems?
For beginners, the Desktop app gives fewer problems today: it installs PyTorch by itself, has the Manager already active and saves a copy before every update. The official ComfyUI website also recommends the app for regular users. The portable version is still useful if you want everything in one folder, maybe on an external drive, and for AMD and Intel cards.
Why does ComfyUI say a model is missing when I've already downloaded it?
Almost always the file is in the wrong folder or has a different name from the one written in the workflow. Check the folder table in this guide, press R in the interface to refresh the list and pick the file from the node's drop-down menu.
Does ComfyUI work without an NVIDIA card?
Yes: there are packages for AMD (RX 6000 and later) and Intel Arc, and on the Mac it works with Apple Silicon chips. Without a dedicated graphics card ComfyUI still starts, but using only the processor it's very slow, especially with recent models. If your computer is not enough, our comparison AI images: local or online? also covers the online alternatives.
Do I need to install Python or Git to use ComfyUI?
No. Both the portable version and the Desktop app already include their own Python. Git isn't needed for normal use: you may only need it to install some nodes directly from a GitHub address.










