MobileSAM_lite

An unofficial Python package for MobileSAM and MobileSAMv2 runtime that adds support for lighter encoder models not available in the original implementation.

This package vendors the runtime code needed for inference:

• mobilesamv2
• tinyvit
• efficientvit
• ultralytics under mobilesam_lite/_vendor/ultralytics

It intentionally does not bundle model checkpoints. Download weights separately and pass the checkpoint path at runtime.

The optional mobilesamv2.promt_mobilesamv2 module now resolves its Ultralytics dependency from the vendored package in mobilesam_lite._vendor.ultralytics.

Install locally

pip install -e .

Install with pypi

pip install mobilesam-lite

Example

import torch

from mobilesam_lite.mobile_sam import SamPredictor, sam_model_registry

model = sam_model_registry["vit_t"]("./weight/mobile_sam.pt")
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
model.eval()

predictor = SamPredictor(model)

Verify an installed wheel

After installing the wheel into a clean environment, run:

python example_inference_mobilesam.py --checkpoint /path/to/mobile_sam.pt

You can also provide a real image:

python example_inference_mobilesam.py --checkpoint /path/to/mobile_sam.pt --image /path/to/image.jpg

The script prints the installed distribution version, the imported package path, and the output tensor shapes from one prediction call.

For the MobileSAMv2 decoder path, use:

python example_inference_mobilesamv2.py \
  --checkpoint /path/to/mobile_sam.pt \
  --prompt-decoder-checkpoint /path/to/Prompt_guided_Mask_Decoder.pt \
  --object-aware-model-checkpoint /path/to/ObjectAwareModel.pt \
  --image /path/to/image.jpg \
  --output-dir wheel_verify_mobilesamv2_output

This script runs the MobileSAMv2 seg-every pipeline with ObjectAwareModel box proposals plus the prompt-guided decoder.

Inputs:

• --checkpoint: image encoder checkpoint
• --prompt-decoder-checkpoint: Prompt_guided_Mask_Decoder.pt
• --object-aware-model-checkpoint: ObjectAwareModel.pt
• --image: optional input image path. If omitted, the script uses a synthetic test image.
• --output-dir: directory for generated visualizations
• Optional tuning args: --encoder-type, --imgsz, --iou, --conf, --retina, --decoder-batch-size, --min-box-area-ratio, --max-box-area-ratio

Outputs:

• Console summary with device, input image shape, detected box count, filtered box count, mask tensor shape, and saved output path
• boxes.png: detected boxes after filtering
• mask_union.png: binary union of all predicted masks
• mask_union_overlay.png: union mask blended over the input image
• mask_overlay.png: per-mask color overlay for the seg-every result

Example assets for the MobileSAMv2 seg-every flow:

Input image:

Output overlay:

Reference: Official MobileSAM repository

https://github.com/chaoningzhang/mobilesam

If you find this repo useful to you please consider click the button below to donate and support my work!