SAM 2.1 Hiera Tiny — ONNX

FP32 ONNX export of facebook/sam2.1-hiera-tiny, split into an image encoder and a prompt encoder/mask decoder for point- and box-prompted image segmentation. The ONNX graphs are mirrored unchanged from the onnx-community export.

Mirrored for use with inference4j, an inference-only AI library for Java.

Original Source

Repository: Meta (ONNX export by onnx-community)
License: apache-2.0

Usage with inference4j

try (Sam2Segmenter segmenter = Sam2Segmenter.builder().build()) {
    SegmentationMask mask = segmenter.segment(
        Path.of("photo.jpg"),
        List.of(new PromptPoint(320, 240, true)));
    ImageIO.write(mask.toImage(), "png", Path.of("mask.png").toFile());
}

Model Details

Property	Value
Architecture	SAM 2.1 Hiera Tiny
Task	Promptable image segmentation
Graphs	`vision_encoder.onnx` and `prompt_encoder_mask_decoder.onnx` (opset 18, FP32)
External data	Each graph loads its matching `.onnx_data` sidecar from the same directory
Image input	`pixel_values`: `[batch, 3, 1024, 1024]` float32, RGB
Image preprocessing	Resize to 1024 x 1024 without preserving aspect ratio; rescale by 1/255; ImageNet mean/std normalization
Prompts	Point coordinates and boxes use the 1024 x 1024 input coordinate space
Mask output	`pred_masks`: `[batch, prompts, 3, 256, 256]` float32 low-resolution logits
Mask selection	Choose one of three candidate masks using `iou_scores`
Refinement	No previous-mask (`mask_input`) input; iterative mask refinement is not supported by this export

License

The original SAM 2.1 model and weights are licensed under the Apache License 2.0. Original model by Meta; ONNX conversion by onnx-community.

Files

The FP32 model remains split into two ONNX graphs and their external-data sidecars:

vision_encoder.onnx + vision_encoder.onnx_data
prompt_encoder_mask_decoder.onnx + prompt_encoder_mask_decoder.onnx_data

All four files must remain in the same directory. The sidecars are deliberately not folded into the graph files.

Prompt Encoding

Foreground point label: 1
Background point label: 0
Padding point label: -10
Box layout: (x1, y1, x2, y2)

Coordinates are expressed in the resized 1024 x 1024 image space. See the source processor_config.json for processor defaults and the inference4j wrapper for conversion from original-image coordinates.

Outputs

The decoder returns three candidate low-resolution mask logits per prompt, their iou_scores, and object_score_logits. Resize the selected logits to the original image size and threshold at zero to obtain a binary mask.

inference4j/sam2.1-hiera-tiny