SAM 2.1 Hiera Tiny, ONNX (fp32 and fp16)

An unchanged copy of the ONNX export of Meta's SAM 2.1 Hiera Tiny that browser-perception runs in production, kept by Imaging for Good so it stays available. Only the four files the library uses are copied, plus the export's configuration files. We did not train, convert or modify this model.

Origin

Model: SAM 2.1 Hiera Tiny by Meta FAIR, Apache-2.0. Code: facebookresearch/sam2. Checkpoint: facebook/sam2.1-hiera-tiny. Paper: Ravi et al., SAM 2: Segment Anything in Images and Videos, arXiv:2408.00714.
ONNX export: by onnx-community, from onnx-community/sam2.1-hiera-tiny-ONNX at commit 814a066640debee5a91e70aa401fb8e17e030503. That repository also has quantized variants (int8, q4, …), not copied here.

Files

File	Bytes	SHA-256
`onnx/vision_encoder_fp16.onnx`	314,925	`7773e79d589dada8e1e630cc7eb18b17b2c5b4faeb1eaab5ebf9df618ecbd50a`
`onnx/vision_encoder_fp16.onnx_data`	67,005,504	`a4dd3759e9b6a476d991fb3493787992e78777e51462695e1bede71ae258103e`
`onnx/vision_encoder.onnx`	354,238	`4f30aacd3aaefbca81a0b7fe4c1fc96345570ea0a6f80ced599493d1b3be2e8c`
`onnx/vision_encoder.onnx_data`	134,084,864	`e83df9866a5afe68ea7f0f721f18f65137fc3acbf0da1c74e946d363e09c69cc`
`onnx/prompt_encoder_mask_decoder_fp16.onnx`	229,799	`00ac7200451f246e8307aa1a1648113f47dc45625738406403a0e6b86c69fc55`
`onnx/prompt_encoder_mask_decoder_fp16.onnx_data`	10,454,016	`027b457a53f5bd76bb401c59555448451d04b4eab6d2f6a6758406b40133ee2c`
`onnx/prompt_encoder_mask_decoder.onnx`	213,114	`874414704c5d686db7d206a35f6e15d26563d50c8c4468fccc6739bd7e491dcf`
`onnx/prompt_encoder_mask_decoder.onnx_data`	20,958,208	`e9874d900dd4134ed60eab1e97910327c2419e0b2954485d8fd6e7f1a1470f47`
Also `config.json`, `preprocessor_config.json` and `processor_config.json`, unchanged.
Each `.onnx` file keeps its weights in the `.onnx_data` file next to it.

Contract

vision_encoder: pixel_values [1, 3, 1024, 1024] float32 → image_embeddings.0 [1, 32, 256, 256], .1 [1, 64, 128, 128], .2 [1, 256, 64, 64].
prompt_encoder_mask_decoder: input_points, input_labels, input_boxes and the three embeddings → iou_scores [1, 1, 3], pred_masks [1, 1, 3, 256, 256] (logits), object_score_logits.
The fp16 files need a GPU with shader-f16 on WebGPU; the fp32 files run anywhere.