steveyin-fresh / osam

Get up and running with SAM, Efficient-SAM, and other segment-anything models locally.

Osam

Get up and running with segment-anything models locally.

Osam (/oʊˈsɑm/) is a tool to run open-source segment-anything models locally (inspired by Ollama).

Osam provides:

• Segment-Anything Models - original SAM, Efficient-SAM;
• Local APIs - CLI & Python & HTTP interface;
• Customization - Host custom vision models.

Installation

Pip

pip install osam

Quickstart

To run with Efficient-SAM:

osam run efficient-sam --image <image_file>

Model library

Here are models that can be downloaded:

Model	Parameters	Size	Download
SAM 91M	91M	100MB	osam run sam:91m
SAM 308M	308M	320MB	osam run sam:308m
SAM 636M	636M	630MB	osam run sam
Efficient-SAM 10M	10M	40MB	osam run efficient-sam:10m
Efficient-SAM 25M	25M	100MB	osam run efficient-sam

PS. sam, efficient-sam is equivalent to sam:latest, efficient-sam:latest.

Usage

CLI

# Run a model with an image
osam run efficient-sam --image examples/_images/dogs.jpg > output.png

# Get a JSON output
osam run efficient-sam --image examples/_images/dogs.jpg --json
# {"model": "efficient-sam", "mask": "..."}

# Give a prompt
osam run efficient-sam --image examples/_images/dogs.jpg \
  --prompt '{"points": [[1439, 504], [1439, 1289]], "point_labels": [1, 1]}' > output.png

Input and output images ('dogs.jpg', 'output.png').

Python

import osam.apis
import osam.types

request = osam.types.GenerateRequest(
    model="efficient-sam",
    image=np.asarray(PIL.Image.open("examples/_images/dogs.jpg")),
    prompt=osam.types.Prompt(points=[[1439, 504], [1439, 1289]], point_labels=[1, 1]),
)
response = osam.apis.generate(request=request)
PIL.Image.fromarray(response.mask).save("mask.png")

Input and output images ('dogs.jpg', 'mask.png').

HTTP

# Get up the server
osam serve

# POST request
curl 127.0.0.1:11368/api/generate -X POST \
  -H "Content-Type: application/json" \
  -d "{\"model\": \"efficient-sam\", \"image\": \"$(cat examples/_images/dogs.jpg | base64)\"}" \
  | jq -r .mask | base64 --decode > mask.png

License

MIT