Deploying ComfyUI on Mac M2: A Record of Pitfalls

Environment

MacBook Pro M2 Max, 32GB RAM
macOS 15 Sequoia
Python 3.11

Step 1: Install ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Pitfall 1: PyTorch MPS Error

The following appears on first launch:

RuntimeError: MPS backend out of memory

Cause: MPS memory management is less mature than CUDA, so large models are prone to OOM.

Solution: Add --lowvram to the startup command:

python main.py --lowvram

Or limit the MPS cache:

export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.5

Pitfall 2: Model Paths

ComfyUI reads from models/ by default. When placing models, pay attention to the subdirectory structure:

ComfyUI/
└── models/
    ├── checkpoints/     # SD base model (.safetensors)
    ├── vae/             # VAE model
    ├── clip/            # CLIP model
    ├── loras/           # LoRA weights
    └── controlnet/      # ControlNet model

Pitfall 3: Slow HuggingFace Downloads

Use huggingface-cli with a mirror to speed things up:

pip install huggingface_hub
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download stabilityai/stable-diffusion-xl-base-1.0 --local-dir models/checkpoints/

Final Configuration

My startup script start.sh:

#!/bin/bash
source venv/bin/activate
export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.3
python main.py --lowvram --listen 0.0.0.0 --port 8188

Performance Data

Model	Resolution	Time	Memory
SD 1.5	512×512	~15s	~6GB
SDXL	1024×1024	~45s	~14GB
SDXL + LoRA	1024×1024	~50s	~16GB

Real-World Use Cases

Scenario 1: E-commerce Product Image Generation You need to generate product images with different backgrounds for 100 SKUs. Configure a ComfyUI workflow: input a product image on a white background → ControlNet extracts the outline → prompt describes the scene ("product on a marble desk, soft lighting, studio quality") → batch generation. One SDXL image takes about 45 seconds, and 100 images take about 75 minutes, saving substantial costs compared with using cloud services.

Scenario 2: UI Design Inspiration Board Designers need to quickly generate mockups in different styles. Using ComfyUI + LoRA models (such as midjourney-style), they can enter a short prompt to get high-quality reference images. Compared with manual search, efficiency improves by more than 10 times.

Advanced Optimization: Using ComfyUI-Manager to Manage Plugins

# Install ComfyUI-Manager
cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager.git
pip install -r ComfyUI-Manager/requirements.txt

# Restart ComfyUI, then install commonly used nodes with one click in Manager:
# - ControlNet preprocessor
# - IPAdapter (style transfer)
# - WAS Node Suite (image processing toolkit)
# - ComfyUI-Impact-Pack (automatic detection)

Performance Tuning: Custom Startup Parameters

#!/bin/bash
# start-optimized.sh
source venv/bin/activate
# MPS Memory Control
export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.3
# Reduce Memory Usage with fp16
export PYTORCH_ENABLE_MPS_FALLBACK=1
python main.py \
  --lowvram \
  --fp16-vae \
  --force-fp16 \
  --listen 0.0.0.0 \
  --port 8188 \
  --front-end-version Comfy-Org/ComfyUI_frontend@latest

Conclusion

Running ComfyUI on a Mac M2 is feasible, but it requires reasonable memory management. Running a 7B LLM and SD at the same time will exhaust RAM, so using them separately is recommended. The MPS ecosystem is still improving, but it is already sufficient for daily use.