Successfully runs CogVideoX-5B (10GB model) for text-to-video generation on Mac M4 Pro with 64GB RAM using Apple's Metal Performance Shaders (MPS).

• ✅ Fixed MPS autocast incompatibility - The main blocker for Mac users
• ✅ Resolved node registration errors - Custom patches for ComfyUI
• ✅ Found optimal memory limits - 30 frames max for stable generation
• ✅ 4-second videos in ~18 minutes - No cloud GPUs needed
• ✅ 99% GPU utilization - Full Apple Silicon power

• RAM: 64GB (critical - 32GB insufficient)
• Storage: 50GB free (models + workspace)
• Chip: M4 Pro or better

During generation: 99% GPU usage
Memory bandwidth: 273 GB/s fully utilized
Power draw: ~50W sustained
Thermals: Cool (< 60°C)

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

cd custom_nodes
git clone https://github.com/kijai/ComfyUI-CogVideoXWrapper
cd ComfyUI-CogVideoXWrapper
pip install -r requirements.txt

Critical: The default nodes don't work on Mac. Apply our fixes:

# Backup original
cp nodes.py nodes.py.backup

# Download and apply our fixed version
curl -o nodes.py https://raw.githubusercontent.com/nicedreamzapp/CogVideoX-Mac-Setup/main/fixes/nodes_fixed.py

Find line ~755 in nodes.py and replace:

# OLD (breaks on Mac):
autocast_context = torch.autocast(
    mm.get_autocast_device(device), dtype=dtype
) if any(q in model["quantization"] for q in ("e4m3fn", "GGUF")) else nullcontext()

# NEW (Mac compatible):
# MPS-safe autocast with fallback
if str(device.type) == "mps":
    autocast_context = nullcontext()
else:
    autocast_context = torch.autocast(
        mm.get_autocast_device(device), dtype=dtype
    ) if any(q in model["quantization"] for q in ("e4m3fn", "GGUF")) else nullcontext()

Around line ~684, verify these default values exist:

else:
    # Default dimensions for text-to-video
    H = 60  # 480 / 8
    W = 90  # 720 / 8
    B = 1
    T = num_frames
    C = 16
    latents = None

# T5 Text Encoder (9GB)
cd ../../models/clip
wget https://huggingface.co/mcmonkey/google_t5-v1_1-xxl_encoderonly/resolve/main/t5xxl_fp16.safetensors

# CogVideoX-5B will auto-download on first run (10GB)

cd ~/Desktop/ComfyUI
export PYTORCH_ENABLE_MPS_FALLBACK=1
export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0
python main.py --force-fp16 --lowvram

1. Open http://127.0.0.1:8188
2. Download our workflow
3. Drag into ComfyUI browser
4. Adjust settings (see below)
5. Queue Prompt!

For Maximum Success Rate:

Empty Latent Image:
  width: 512
  height: 384
  batch_size: 1

CogVideo Sampler:
  num_frames: 30      # CRITICAL: Don't exceed!
  steps: 20-30        # Higher = better quality but slower
  cfg: 6.0-6.5
  scheduler: CogVideoXDDIM
  
Load CLIP:
  clip_name: t5xxl_fp16.safetensors
  type: sd3

• Max frames: 30 (31+ causes white screen corruption)
• Max steps: 30 (50+ exhausts memory during VAE decode)
• Resolution: 512x384 (higher resolutions untested/risky)

Symptoms: Generation completes 100% but video is white with artifacts

Causes:

• Too many frames (>30)
• Too many steps (>30)
• Memory exhaustion during VAE decode

Solutions:

1. Reduce to 25-30 frames
2. Use 20-30 steps maximum
3. Enable VAE tiling in CogVideo Decode node
4. Restart ComfyUI to clear memory

Error: Failed to allocate private MTLBuffer for size XXXGB

Solutions:

• Reduce frames to 25
• Reduce steps to 20
• Close all other applications
• Restart Mac to clear unified memory

Error: Cannot import CogVideoXWrapper module

Solutions:

• Check indentation in nodes.py (line 674 must have 8 spaces)
• Verify fixes were applied correctly
• Restore from nodes.py.backup and reapply fixes

CogVideoX-Mac-Setup/
├── README.md
├── fixes/
│   ├── nodes_original.py      # Backup of original
│   └── nodes_fixed.py          # Mac-compatible version
├── workflows/
│   └── cogvideox_1_0_5b_T2V_02.json
└── examples/
    ├── videos/          # (hosted on GitHub)
    └── screenshots/

What We Fixed:

1. MPS Autocast Bug: PyTorch's torch.autocast() doesn't support MPS device type, causing RuntimeError. Wrapped in device check with nullcontext() fallback.

2. H/W Variable Bug: Text-to-video mode didn't initialize height/width when no input image present, causing UnboundLocalError. Added default values based on model expectations.

3. Memory Limits: Through testing, discovered 30-frame hard limit on M4 Pro 64GB due to VAE decode memory requirements (~46GB buffer allocation at 49 frames).

Found this helpful? Consider:

• ⭐ Star this repo
• 🐛 Report issues you encounter
• 💡 Share your successful configurations
• 🎥 Submit example videos

• ComfyUI - The UI framework
• CogVideoX - Original model
• ComfyUI-CogVideoXWrapper - The wrapper we fixed

Verdict: Mac is perfect for learning/experimenting. RTX for production.

This guide and fixes are MIT licensed. Original CogVideoX and ComfyUI have their own licenses.

• kijai for the ComfyUI-CogVideoXWrapper
• THUDM for CogVideoX model
• comfyanonymous for ComfyUI
• Apple for M4 Pro silicon that makes this possible