Stable Diffusion on MLX (SD 1.4 / SD 1.5)

Run Stable Diffusion on Apple Silicon using MLX. This repo focuses on SD 1.x models, with a fast, minimal pipeline and a simple CLI. By default it loads NovelAI/nai-anime-v2 UNet + VAE with the SD 1.4 text stack, but you can also run most SD1.5 checkpoints from CivitAI with a single file.

Highlights

• SD 1.4 + SD 1.5 model support on MLX
• Works great on older Macs
• CivitAI .safetensors SD1.5 checkpoint loading (UNet + text encoder)
• Optional custom VAE from checkpoint
• Simple environment-variable driven prompts

---

Quickstart

Install deps and generate your first image:

pip install -r requirements.txt

SD_PROMPT="flowers, flower field, sunset, no humans" \
SD_NEGATIVE_PROMPT="humans" \
SD_SEED=56 \
python3 t2i_sd.py

Output defaults to output.png (use -o to change the filename).

---

Use a CivitAI SD1.5 Checkpoint

Most SD1.5 models from CivitAI are now supported. Download a SD1.5 .safetensors file and run:

python3 t2i_sd.py --ckpt /path/to/model.safetensors

If you need a different base config/tokenizer:

python3 t2i_sd.py --ckpt /path/to/model.safetensors --base-model runwayml/stable-diffusion-v1-5

By default the VAE comes from the base model for max compatibility. To use the checkpoint’s VAE:

python3 t2i_sd.py --ckpt /path/to/model.safetensors --use-ckpt-vae

---

CLI Options

• -o output.png: output filename
• -n, --no-novelai: use CompVis/stable-diffusion-v1-4
• --ckpt /path/to/model.safetensors: load SD1.5 checkpoint
• --base-model runwayml/stable-diffusion-v1-5: base configs/tokenizer
• --use-ckpt-vae: use the checkpoint’s VAE

---

Prompt Controls (Environment Variables)

• SD_PROMPT: positive prompt
• SD_NEGATIVE_PROMPT: negative prompt
• SD_CFG: classifier-free guidance (default 7.5)
• SD_STEPS: denoising steps (default 50)
• SD_SEED: seed (0 picks a random one)

Example:

SD_PROMPT="1girl, cinematic lighting" \
SD_NEGATIVE_PROMPT="ugly, disfigured" \
SD_CFG=7.0 SD_STEPS=28 SD_SEED=123 \
python3 t2i_sd.py --ckpt /path/to/model.safetensors

---

Example Outputs

With SD_STEPS around 28 you can get strong results quickly; 50 steps is the default for higher detail.

---

Notes

• SD 1.4 is especially strong for stylized artwork.
• SD 1.5 shares the same CLIP ViT-L/14 text encoder, which this repo uses.
• SD2/SDXL checkpoints are not supported.