kornia/sam

Pretrained weights for Segment Anything (SAM), used by kornia.models.sam.Sam.

SAM predicts object masks from point, box or mask prompts. It pairs a ViT image encoder with a lightweight prompt encoder and mask decoder.

Original repo: facebookresearch/segment-anything

Weights

File	Image encoder
sam_vit_b_01ec64.safetensors	ViT-B
sam_vit_l_0b3195.safetensors	ViT-L
sam_vit_h_4b8939.safetensors	ViT-H

Provenance

Each file holds the tensors of the upstream checkpoint with the same stem, unchanged: the same keys, dtypes, shapes and bits. The kornia maintainers converted them (converter revision 0c3db846b2). They checked that a model loaded from either file has a bitwise-identical state_dict() and gives bitwise-identical image-encoder outputs.

File	Upstream source	Source sha256	sha256
sam_vit_b_01ec64.safetensors	sam_vit_b_01ec64.pth	ec2df62732614e57411cdcf32a23ffdf28910380d03139ee0f4fcbe91eb8c912	fbf4c1ad844a2e9b867d89dc04a61f3ece19c57ebb7d8fcb6da92062a291d839
sam_vit_l_0b3195.safetensors	sam_vit_l_0b3195.pth	3adcc4315b642a4d2101128f611684e8734c41232a17c648ed1693702a49a622	8e087bc0db5cfac2676ae69bb1b846fa1f6cc193fc0beb6be8fb14fc7406be41
sam_vit_h_4b8939.safetensors	sam_vit_h_4b8939.pth	a7bf3b02f3ebf1267aba913ff637d9a2d5c33d3173bb679e46d9f338c26f262e	1424ffbab6fa8218aeae5abb931549dcd2d68725220ed14a5d1b579aa503c3e7

The upstream ViT-B and ViT-L checkpoints store mask_decoder.iou_token.weight and mask_decoder.mask_tokens.weight as views of one shared storage. The safetensors format forbids shared storage, so each is stored as its own copy with identical values; loading a model copies them into separate parameters either way.

License

Apache-2.0. The upstream README states that "The model is licensed under the Apache 2.0 license". See LICENSE, copied from facebookresearch/segment-anything. The SA-1B dataset has its own licence, which does not apply to these weights.

Citation

@article{kirillov2023segany,
    title   = {Segment Anything},
    author  = {Kirillov, Alexander and Mintun, Eric and Ravi, Nikhila and Mao, Hanzi and Rolland, Chloe and
               Gustafson, Laura and Xiao, Tete and Whitehead, Spencer and Berg, Alexander C. and Lo, Wan-Yen and
               Doll{\'a}r, Piotr and Girshick, Ross},
    journal = {arXiv:2304.02643},
    year    = {2023}
}