Stable Diffusion on Mac: Apple Silicon M-Series Setup for Uncensored AI in 2026
How to run Stable Diffusion locally on Mac with Apple Silicon M1, M2, M3, or M4. Full setup guide for uncensored AI image generation on macOS — which UIs work, performance expectations, and model compatibility.
Yes, Macs Run Stable Diffusion
Apple Silicon changed the game. The unified memory architecture means your M1/M2/M3/M4 Mac can load large AI models that would require expensive dedicated VRAM on a PC. It's slower than NVIDIA, but it works — and it's completely uncensored and offline.
If you have a Mac with Apple Silicon and 16+ GB memory, you can generate AI images locally right now.
Mac Performance Expectations
| Mac | Memory | SDXL (1024×1024) | Flux |
|---|---|---|---|
| M1 (base) | 8–16 GB | ~60–90 sec | Too slow / OOM on 8 GB |
| M1 Pro/Max | 16–64 GB | ~30–50 sec | Works on 32+ GB |
| M2 Pro/Max | 16–96 GB | ~20–40 sec | Comfortable on 32+ GB |
| M3 Pro/Max | 18–128 GB | ~15–30 sec | Good on 36+ GB |
| M4 Pro/Max | 24–128 GB | ~12–25 sec | Best Mac experience |
Times are approximate for 20 steps, DPM++ 2M Karras. Actual performance varies by model and settings.
Best UI for Mac: ComfyUI
ComfyUI has the best macOS support in 2026:
- Install Python 3.11+ via Homebrew:
brew install python@3.11 - Clone ComfyUI:
git clone https://github.com/comfyanonymous/ComfyUI - Install dependencies:
pip install -r requirements.txt - Download a model (Juggernaut XL) into
models/checkpoints/ - Launch:
python main.py --force-fp16
The --force-fp16 flag is important on Mac — it uses half-precision which is faster and uses less memory on MPS.
Alternative: Draw Things (Mac-native app)
Draw Things is a free, native macOS/iOS app built specifically for Apple Silicon. It's the easiest way to get started — no terminal required. Supports SDXL, ControlNet, LoRAs, and runs offline.
Mac-Specific Tips
- More memory = bigger models. On Mac, unified memory IS your VRAM. 16 GB handles SDXL, 32+ GB opens up Flux and larger models.
- Close other apps while generating. Safari, Chrome, and Xcode all compete for the same memory pool.
- Use FP16 always. FP32 is 2× slower and uses 2× memory on MPS with minimal quality difference.
- MPS vs CPU: Always ensure MPS acceleration is active (it should be automatic). If you see "Using device: cpu" in logs, something is wrong.
- Batch size 1 only. Higher batch sizes on Mac MPS often cause errors or are slower than sequential single-image generation.
Mac works — but it's slower and harder to set up. If you also have a Windows PC with an NVIDIA GPU (even a laptop), LocalForge AI will get you generating 3–5× faster with zero configuration. One download, one click, done.
FAQ
Can I run Stable Diffusion on an Intel Mac?
Technically yes, but it runs on CPU only — painfully slow (5–10+ minutes per image). Not practically usable. Apple Silicon is required for decent performance.
Does ControlNet work on Mac?
Yes, in ComfyUI. Most ControlNet types (OpenPose, Canny, Depth) work with MPS acceleration. Some preprocessors may fall back to CPU but still function.
Can I train LoRAs on Mac?
Experimental support exists via Kohya_ss and ml-stable-diffusion. Training is very slow compared to NVIDIA — expect 3–5× longer training times. Feasible for small datasets, impractical for large ones.