grabbou/EfficientSAM-Ti-CoreAI

EfficientSAM-Ti for Core AI

EfficientSAM ViT-Tiny (10M parameters) exported to Apple Core AI, unmodified, with the apple/coreai-models recipe. It runs on iOS 27 and macOS 27 with the ImageSegmenter from the CoreAISegmentation product of coreai-models.

It segments the object at a point (a click) or in a box, one mask per prompt.

Files

efficient-sam/ is a segmenter bundle: metadata.json and the exported model, float32, static shape, up to two points per prompt (39 MB).

Use

With @react-native-ai/core-ai, download the folder and pass its path:

import { CoreAI } from '@react-native-ai/core-ai'

const result = await CoreAI.runTask(
  'segmentation',
  { id: 'efficient-sam', sourceType: 'file', sourceUri: efficientSAMBundlePath },
  { imageUri, data: { box: { x: 10, y: 50, width: 300, height: 400 } } },
  {}
)

In Swift:

let segmenter = try await ImageSegmenter(resourcesAt: "efficient-sam")
let response = try await segmenter.segment(image: cgImage, pointQuery: query)

Export

From apple/coreai-models at 09ff139, with coreai-core 1.0.0b3:

cd models/efficient-sam
uv run export.py

License

Apache-2.0, as the original model. EfficientSAM is by Y. Xiong et al., EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything.