SAMExporter 0.5.0 expands portable ONNX support and ships reviewed real-model evidence.
Highlights:
- EfficientSAM-Ti/S and complete MobileSAM encoder/decoder support
- latest pinned official SAM, SAM2/SAM2.1, and SAM3 sources
- corrected SAM1 aspect ratios/colors and SAM2 original-size mask quality
- SAM3 raw-head export, official-style mask NMS, prompt-focused selection, text examples, and flexible instance limits
- provider selection and extras for CPU, CUDA/TensorRT, OpenVINO, DirectML, QNN, CoreML, and other ONNX Runtime builds
- MIT-licensed packaged CLIP tokenizer with no forced CPU runtime dependency
- public validated models: https://huggingface.co/nrl-ai/samexporter-onnx-models
- full-resolution outputs and logs: https://github.com/vietanhdev/samexporter/tree/v0.5.0/visual_results
Validation: 42 tests pass on Python 3.11 and 3.13; real SAM ViT-B, MobileSAM, EfficientSAM-Ti, SAM2.1 Tiny, and SAM3 outputs were visually reviewed.