Skip to main content
This guide covers frequently encountered issues and their solutions when running ComfyUI.

​
Installation Issues

This error occurs when PyTorch is installed without CUDA support.Solution:
  1. Uninstall the current PyTorch installation: 
    pip uninstall torch
  2. Reinstall PyTorch with CUDA support: 
    pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
For AMD GPUs, use the ROCm version instead:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.1
If you have trouble extracting the portable Windows version, the file may be blocked by Windows.Solution:
  1. Right-click the downloaded .7z file
  2. Select Properties
  3. Check the Unblock checkbox at the bottom
  4. Click OK and try extracting again
Custom nodes can sometimes conflict or fail to load properly.Solution:Disable all custom nodes to isolate the issue:
python main.py --disable-all-custom-nodes
To enable only specific nodes:
python main.py --disable-all-custom-nodes --whitelist-custom-nodes "node_name_1,node_name_2"
This helps identify which custom node is causing problems.

​
Memory Issues

OOM errors occur when your GPU runs out of VRAM during generation.Solution:Use low VRAM mode to reduce memory usage:
python main.py --lowvram
For extremely limited VRAM (< 2GB):
python main.py --novram
For systems without a GPU:
python main.py --cpu
Note: CPU mode is significantly slower than GPU mode.
A known bug in xformers version 0.0.18 causes black images when generating high-resolution outputs.Solution:Downgrade or upgrade xformers to a different version:
pip install xformers!=0.0.18
Or disable xformers entirely:
python main.py --disable-xformers
If you see a warning about cudaMallocAsync compatibility with your GPU.Solution:Disable CUDA malloc:
python main.py --disable-cuda-malloc
This is particularly relevant for older GPUs that don’t support async memory allocation.

​
Runtime Errors

ComfyUI can’t find your checkpoint files.Solution:
  1. Ensure models are in the correct directory:
    • Checkpoints: models/checkpoints
    • VAE: models/vae
    • LoRAs: models/loras
    • Embeddings: models/embeddings
  2. Configure extra model paths by copying extra_model_paths.yaml.example to extra_model_paths.yaml and editing it: 
    cp extra_model_paths.yaml.example extra_model_paths.yaml
  3. Edit extra_model_paths.yaml with your model paths
  4. Restart ComfyUI
Latent previews may not be enabled or TAESD models are missing.Solution:
  1. Enable preview method: 
    python main.py --preview-method auto
  2. For high-quality previews, download TAESD decoder files:
    • taesd_decoder.pth
    • taesdxl_decoder.pth
    • taesd3_decoder.pth
    • taef1_decoder.pth
  3. Place them in models/vae_approx directory
  4. Launch with TAESD preview: 
    python main.py --preview-method taesd
The execution may be stuck or waiting for resources.Solution:
  1. Check if a node has invalid inputs or missing connections
  2. Cancel the current execution with Ctrl + Alt + Enter
  3. If the issue persists, restart ComfyUI
  4. Check the console for error messages
Note: Only parts of the graph with all correct inputs will execute. Incomplete workflows will not run.

​
Platform-Specific Issues

macOS versions 14.5 and later have a known bug causing black images.Solution:Force upcast attention (automatically enabled on macOS 14.5+):
python main.py --force-upcast-attention
This forces computations to use higher precision, avoiding the black image bug.
Windows reserves additional VRAM for shared memory management.Solution:ComfyUI automatically reserves 600MB on Windows (vs 400MB on Linux). If you still experience issues:
python main.py --reserve-vram 1.0
This reserves 1GB of VRAM for the system. Adjust the value based on your needs.
AMD GPUs may need specific environment variables or settings.Solution:For unsupported AMD cards, override the architecture version:RDNA2 (6700, 6600):
HSA_OVERRIDE_GFX_VERSION=10.3.0 python main.py
RDNA3 (7600):
HSA_OVERRIDE_GFX_VERSION=11.0.0 python main.py
Enable experimental memory efficient attention:
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention
Enable TunableOp (slow first run, faster subsequent runs):
PYTORCH_TUNABLEOP_ENABLED=1 python main.py

​
Database and Assets

The asset database may fail to initialize if dependencies are missing.Solution:
  1. Ensure all requirements are installed: 
    pip install -r requirements.txt
  2. If the error persists, disable automatic asset scanning: 
    python main.py --disable-assets-autoscan
  3. Report persistent database issues as they may indicate a dependency problem

​
Getting Help

If you’ve tried the solutions above and still have issues:
  1. Discord: Join the ComfyUI Discord and try the #help or #feedback channels
  2. Matrix: Join the ComfyUI Matrix space at #comfyui_space:matrix.org
  3. GitHub Issues: Report bugs at the ComfyUI GitHub repository
When reporting issues, include:
  • Your operating system
  • GPU model and VRAM
  • Python version
  • PyTorch version
  • The full error message from the console
  • Steps to reproduce the issue

Build docs developers (and LLMs) love

Get started for freeTalk to us