Installation

Basic install

pip install strands-sana

This pulls in diffusers, torch, transformers, accelerate, pyyaml.

With GPU acceleration

You'll want a CUDA-compatible torch wheel for non-trivial inference:

# CUDA 12.1
pip install torch>=2.4 --extra-index-url https://download.pytorch.org/whl/cu121

pip install strands-sana

On Apple Silicon, MPS works out of the box — no extra wheel needed.

Optional extras

pip install 'strands-sana[hf]'           # HF Hub + safetensors
pip install 'strands-sana[video]'        # imageio + imageio-ffmpeg (MP4 export)
pip install 'strands-sana[lora]'         # peft for LoRA support
pip install 'strands-sana[quantization]' # bitsandbytes + optimum-quanto
pip install 'strands-sana[dev]'          # pytest, ruff, black, mypy
pip install 'strands-sana[docs]'         # mkdocs-material

With training support

Training tools shell out to upstream NVlabs/Sana scripts. Either:

git clone https://github.com/NVlabs/Sana.git
# Tools auto-detect ./Sana

…or set SANA_ROOT:

export SANA_ROOT=/path/to/Sana

Verify install

import strands_sana
print(strands_sana.__version__)

from strands_sana import sana_load_model
print(sana_load_model(model="list"))

You should see 15 models registered.

Hardware requirements

Model	VRAM (default)	VRAM (mode='low')
Sana 0.6B	4 GB	2 GB
Sana 1.6B / 1.5 1.6B	8 GB	4 GB
Sana 1.5 4.8B	16 GB	8 GB
Sana 1.6B @ 4K	22 GB	8 GB
SANA-Video 2B	12 GB	6 GB
LongSANA	20 GB	10 GB

CPU-only works for tiny experiments but is slow (~5 min/image).