jessolmstead/greycard-sam3

greycard SAM 3

The Parts masks for greycard, a RAW development engine and editor: the iris, teeth, lips, hair, skin, a person's top and the like, each a mask a look can go on. The editor downloads these files on first use of a Parts mask; nothing here needs to be fetched by hand.

This is SAM 3 (Meta, "SAM 3: Segment Anything with Concepts", 2025), converted to ONNX by greycard from Meta's checkpoint sam3.pt at facebook/sam3. It is a derivative of the SAM Materials and is distributed under the SAM License, a copy of which is in LICENSE. greycard is not affiliated with Meta.

Files

• sam3_image_encoder_fp16.onnx (908,762,234 bytes, sha256 dd4733a6882264d0ab00599acd6ece3dd700f9b61b7976369b956ce41d88c23d): the image encoder, its weights and arithmetic converted to fp16.
• sam3_decoder.onnx (129,854,990 bytes, sha256 7c205232b0698d532731fd5fa206cdea694717de69c0d8e88997a14027425564): the decoder, returning the sixteen best queries ranked, with the presence score apart.
• sam3_presets.bin (524,962 bytes, sha256 a1c205ef7e4374a6a9072928339d957eb78f78efa01b5cf170fe0f6ff92d6917): the language encoder's output for the sixteen phrases the editor's menu asks for, computed once, so no text encoder ships. sam3_presets.json lists them.

What was changed

• Traced to ONNX with Meta's model code (at an ONNX-export branch that replaces the complex-valued rotary encoding with cosine and sine) as three graphs: image encoder, language encoder, decoder.
• The image encoder converted to fp16; its masks match the fp32 export's at a median IoU of 0.9996 on the trial frames.
• The decoder made to return the sixteen best queries with the presence score as its own output, instead of cutting at 0.5 inside the graph.
• The language encoder is not distributed; the phrases' encodings are.

The weights are otherwise Meta's.

Contract

• Image encoder input image: 3 × 1008 × 1008, uint8, RGB, the picture stretched to the square (bilinear). Outputs vision_pos_enc_0..2 and backbone_fpn_0..2, float32.
• Decoder inputs: vision_pos_enc_2, backbone_fpn_0..2, language_mask (bool 1 × 32), language_features (float32 32 × 1 × 256), and a box prompt masked off (box_coords zeros 1 × 1 × 4, box_labels ones int64 1 × 1, box_masks true 1 × 1). Outputs boxes (16 × 4, normalized x0 y0 x1 y1), scores (16, descending: each query's score times the presence score), presence (1), masks (16 × 1 × 288 × 288, sigmoid).
• Presets: little-endian; a u32 count, then for each phrase a u16 byte length and the UTF-8 phrase, 32 mask bytes (0 or 1) and 32 × 256 float32.

Reproducing the files

tools/ai/sam3_trial/ in the greycard repository (export.py, fp16.py, table.py, with the pinned versions in requirements-export.txt; the recipe is in its README) gives these exact bytes from sam3.pt.

License

The SAM License, in LICENSE. By using or distributing these files you agree to its terms, including its rules on trade controls and its requirement to comply with applicable privacy and data protection laws. greycard itself is GPL-3.0-or-later: https://github.com/jessolmstead/greycard.