pixano-inference-sam

The SAM2 models (image segmentation and video tracking) for Pixano Inference, distributed as a standalone plugin package — a first-party example of the "model store" pattern. The server discovers it automatically through the pixano_inference.models entry point.

Install

pip install pixano-inference[sam]   # or: pip install pixano-inference-sam
# the upstream SAM2 library is only on git, so install it too:
pip install "sam-2 @ git+https://github.com/facebookresearch/sam2.git"

The first command pulls in the Pixano Inference core and PyTorch; the second installs Facebook's sam-2 library (not available on PyPI).

From the repository (uv also installs the core from it):

uv pip install "pixano-inference-sam @ git+https://github.com/pixano/pixano-inference#subdirectory=packages/pixano-inference-sam"
# or, from a clone:  uv pip install ./packages/pixano-inference-sam

Models

Use

Reference the model by name in a config file (no import needed once installed):

from pixano_inference.configs import DeploymentConfig, ModelConfig
from pixano_inference_sam import Sam2ImageParams

models = [
    ModelConfig(
        name="sam2-image",
        model_class="Sam2ImageModel",
        model_params=Sam2ImageParams(path="facebook/sam2-hiera-base-plus"),
        deployment=DeploymentConfig(num_gpus=1, num_cpus=2),
    ),
]

Development

The package is self-contained, with its own lock file:

cd packages/pixano-inference-sam
uv sync          # the dev group includes the git-only sam-2 library
uv run pytest