sana-cli

Generate images from a text prompt using SANA Sprint, a fast 1.6B-parameter diffusion model.

Example

Prompt: a mouse eating cheese

Requirements

• Python 3.14+
• uv
• Apple Silicon Mac or Linux with CUDA GPU (device auto-detected: CUDA → MPS → CPU)
• Model weights downloaded locally (Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers)

Usage

On a new machine, download the model weights once:

uv run main.py --download

Then generate images (subsequent runs stay fully offline):

uv run main.py 'a mouse eating cheese'       # defaults to square
uv run main.py 'a mouse eating cheese' -s    # square    (1024x1024)
uv run main.py 'a mouse eating cheese' -p    # portrait  (768x1024)
uv run main.py 'a mouse eating cheese' -l    # landscape (1024x768)
uv run main.py 'a mouse eating cheese' -pl   # portrait & landscape
uv run main.py 'a mouse eating cheese' -spl  # all three sizes

Output files are saved to output/ as output_<timestamp>_<size>.png.

Dependencies

Package	Purpose
torch	Tensor operations and device support (CUDA/MPS/CPU)
diffusers	SANA Sprint pipeline
transformers	Text encoder
accelerate	Optimized model loading