SAM 3, ONNX export for PIXL

SAM 3 by Meta. This repository holds PIXL's ONNX export of Meta's SAM 3 image model: concept prompts ("red car" finds every instance) and exemplar boxes. It is split into three graphs, run by the PIXL engine's segment_concept (@xuckless/pixl-engine).

This is a derivative of the SAM Materials. It is distributed under the SAM License (LICENSE), the same terms as Meta's release, and your use is bound by that agreement, trade controls included. It is not an official Meta release.

Files

File Inputs โ†’ outputs Bytes SHA-256
sam3_image_encoder_fp16.onnx image [1,3,1008,1008] โ†’ the detector's FPN and positional encodings 908,717,454 709e68c56ba6c1e40c0cfa3381b06b29fc358b1281c706e02c1ff45d71dead8c
sam3_text_encoder_fp16.onnx tokens int64 [1,32] โ†’ language_features, language_padding, language_embeds 706,988,044 2145c1a55137ea64076b441b48a79feaeebdcc6932154c9f702e54702a96eab9
sam3_decoder.onnx the encoders' outputs + 8 exemplar slots and their mask โ†’ pred_logits, pred_boxes, pred_masks, presence_logit 129,868,227 f6749e7b7a558ab3b16b0ba75075ebe7086003d86d129f8fa831dea727f01a28

FP16 storage: both encoders store their weights in float16 and compute in float32. On a test photograph their masks match the float32 export exactly (IoU 1.0000, scores equal to 3 places).

The tokenizer is CLIP's BPE merges (bpe_simple_vocab_16e6.txt.gz), from facebookresearch/sam3.

How it was made

  • Source: facebook/sam3 sam3.pt (3,450,062,241 B, sha256 9999e2341ceef5e136daa386eecb55cb414446a00ac2b55eb2dfd2f7c3cf8c9e), code at facebookresearch/sam3 2345a4ad109ac29c569da749c91d84f10dc08c40.
  • Export: tests/models/export/export_sam3.py in xuckless/pixl-engine, run with torch 2.14.0+cpu, onnx 1.23.0 and onnxruntime 1.30.0. It is byte-reproducible.
  • Checked against PyTorch on the same inputs:
    • image encoder within 1.6e-4;
    • text encoder within 1.1e-5;
    • decoder within 4.5e-3 on mask logits up to 69.
  • On a real photograph: "road" gives presence 0.943 and mask IoU 0.970 against Meta's own Sam3Processor.

Attribution

SAM 3 ยฉ Meta Platforms, Inc. Licensed under the SAM License. Exported by PIXL (xuckless).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for xuckless/pixl-sam3-onnx

Base model

facebook/sam3
Quantized
(28)
this model