Contents

ComfyUI FAQ: Common Errors and How to Fix Them

GPU memory shortage: CUDA out of memory / allocation failed

Try these steps in order:

Lower the resolution or batch_size.
Use an FP8 or GGUF-quantized model.
For large images, use VAE Decode (Tiled) for VAE decoding.
Close other programs that use GPU memory, such as browsers, games, and other AI tools.
Add --lowvram when starting ComfyUI.
Click “Unload models” in the interface (its location varies by version), or restart ComfyUI.

Nodes have a red outline

When you run the workflow, an error message identifies the node and input with a problem:

Message	Cause
Required input is missing	A required input is not connected
Value not in list	The selected model file does not exist (for example, you loaded someone else’s workflow but don’t have that model). Select the model again or download it
Node title appears red and is unrecognized	A custom node is missing. Use Manager’s Install Missing Custom Nodes

A model you just downloaded cannot be found

Make sure it is in the correct folder (see Model downloads and folders).
Press R to refresh, or restart ComfyUI.
Make sure the file downloaded completely by checking its size.

Generated images are completely black

VAE precision issue: Try the VAE made for the model, or use a different VAE.
You may be using an incompatible combination, such as a model and VAE from different model families.
Some GPUs can produce black images with certain precision settings. Check the warning messages in the terminal.

Poor or strange image quality

Symptom	Check
Blurry, little detail	Too few steps; resolution is too low
Oversaturated, artifacts	CFG is too high; LoRA strength is too high
Duplicate people, strange composition	Resolution is much higher than the model’s recommendation; generate first, then upscale
Does not match the prompt at all	CFG is too low; LoRA / ControlNet is incompatible with the model version
Extra fingers, distorted faces	Model limitation; fix it with inpainting

Generation is very slow

Check the startup messages in the terminal to confirm that ComfyUI is using the GPU (Device: cuda:0 ...).
If the terminal shows a message like loaded partially, the model does not fit in GPU memory and part of it is in system memory. Switch to a smaller model version.
The first run needs to load the model; later runs will be faster.

Torch not compiled with CUDA enabled

You installed the CPU version of PyTorch. Reinstall the GPU version in ComfyUI’s Python environment:

pip uninstall torch torchvision torchaudio
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

(Use the CUDA version recommended in the ComfyUI README.)

ComfyUI will not start after an update

Often, a custom node is incompatible with the new version:

Check the terminal error message to find the name of the node causing the error.
Temporarily move that node out of the custom_nodes folder, then start ComfyUI.
Update the node, or check its GitHub page for a fix.

Issues with Chinese paths or file names

Some custom nodes may fail when model, image, or installation paths contain Chinese characters or special characters. Use an installation path with English characters and no spaces whenever possible.

Dragging an image from a workflow into ComfyUI does nothing

The workflow information in the image has been removed, for example, by compression on a social media or messaging platform. Get the original PNG or JSON file.

Other computers on the local network cannot connect

Add --listen when starting ComfyUI (for example, --listen 0.0.0.0), and make sure the firewall allows port 8188. ComfyUI does not have password protection, so expose it only on a trusted network.

Where can I find more workflows?

Templates built into ComfyUI
Official model pages, which often include sample workflows
Workflows shared on community sites (check whether the required custom nodes are trustworthy)

Updated October 12, 2026