Sana Mac M2 Quickstart

A clean, Mac-first launcher for generating images with Sana using Diffusers + PyTorch MPS.

This wrapper is built for Apple Silicon inference and intentionally avoids CUDA-only setup paths.

Why this repo

• Apple Silicon focused (M1/M2/M3/M4)
• One-command setup and launch flow
• CLI + Gradio UI + benchmark included
• Upstream Sana kept as reference, not default runtime path

Quick Start

chmod +x setup.sh launch.sh smoke_check.sh
./setup.sh
./launch.sh verify
./launch.sh generate "a small robot building a glowing circuit board"
./launch.sh ui

Commands

./launch.sh --help
./launch.sh verify
./launch.sh generate "your prompt"
./launch.sh generate --prompt "your prompt" --steps 8
./launch.sh generate "your prompt" --negative-prompt "blurry, low quality"
./launch.sh ui
./launch.sh benchmark

Each generation now writes a PNG plus a same-name JSON metadata sidecar. Example: sana_m2_output.png and sana_m2_output.json.

Recommended baseline

• Model: Efficient-Large-Model/Sana_600M_512px_diffusers
• Resolution: 512x512
• Steps: 8 for smoke tests, 20 for higher quality
• Dtype: float16 on MPS

If output is unstable, use float32 for numerical stability:

./launch.sh generate "your prompt" --steps 8 --dtype float32

Negative prompts are passed only when supported by the installed SanaPipeline version. If unsupported, generation continues and prints a warning.

Important safety note

Do not run upstream CUDA installer commands on Mac:

cd Sana-main
bash environment_setup.sh
pip install -e .

Use root-level scripts instead.

Validation before release

Run:

./smoke_check.sh

This checks shell syntax, Python compile health, and guards against CUDA-only root dependency drift.

Docs

• Detailed Mac operator guide: README_M2.md
• Upstream warning file: Sana-main/DO_NOT_RUN_ON_MAC_M2.md

License

• Wrapper scripts in this repository are licensed under MIT: LICENSE
• Upstream Sana code under Sana-main/ follows its own upstream license: Sana-main/LICENSE