This is a safetensors-only mirror of the upstream Meta AI Research weights at the pinned commit shown below. The model.safetensors file is byte-identical to upstream; we do not modify weights or configuration. The legacy pytorch_model.bin (pickle format) that upstream ships alongside safetensors has been deliberately removed from this mirror for security hygiene — pickle loads can execute arbitrary code, and transformers auto-prefers safetensors when both are present, so removing it has zero functional impact on downstream users.

The purpose of the mirror is license hygiene: acaua's core promise is that every shipped weight has an auditable, declared Apache-2.0 upstream. Mirroring lets us pin a specific revision so the audit claim stays verifiable even if upstream rewrites history.

Provenance


Upstream repo	`facebook/mask2former-swin-tiny-coco-instance`
Upstream commit SHA	`22c4a2f15dc88149b8b8d9f4d42c54431fbd66f6`
Upstream commit date	2023-09-11
Declared license	Apache-2.0 (upstream YAML frontmatter)
Paper	Cheng et al., "Masked-attention Mask Transformer for Universal Image Segmentation", CVPR 2022, arXiv:2112.01527
Official code	`facebookresearch/Mask2Former` (MIT)
Backbone	Swin-Tiny, pretrained on ImageNet-1k (per upstream model card)
Mirrored on	2026-04-17
Mirrored by	CondadosAI/acaua

Usage via acaua

python

import acaua
model = acaua.Model.from_pretrained("CondadosAI/mask2former_swin_tiny_coco_instance")
results = model.predict("image.jpg")
for r in results:
    print(r.boxes, r.labels, r.scores, r.masks.shape)

Usage via 🤗 Transformers

This mirror is drop-in compatible with the upstream Facebook repo:

python

from transformers import AutoModelForUniversalSegmentation, AutoImageProcessor
model = AutoModelForUniversalSegmentation.from_pretrained(
    "CondadosAI/mask2former_swin_tiny_coco_instance"
)
processor = AutoImageProcessor.from_pretrained(
    "CondadosAI/mask2former_swin_tiny_coco_instance"
)

License and attribution

Redistributed under Apache License 2.0, consistent with the upstream HF model card declaration. The reference implementation at facebookresearch/Mask2Former is MIT-licensed; the weights as distributed by facebook/* on Hugging Face are declared Apache-2.0.

See NOTICE for required attribution to upstream contributors (Meta AI Research / FAIR, Mask2Former authors, Swin Transformer authors).

Citation

bibtex

@inproceedings{cheng2022mask2former,
  title={Masked-attention Mask Transformer for Universal Image Segmentation},
  author={Cheng, Bowen and Misra, Ishan and Schwing, Alexander G and Kirillov, Alexander and Girdhar, Rohit},
  booktitle={CVPR},
  year={2022}
}

@inproceedings{liu2021swin,
  title={Swin Transformer: Hierarchical Vision Transformer using Shifted Windows},
  author={Liu, Ze and Lin, Yutong and Cao, Yue and Hu, Han and Wei, Yixuan and Zhang, Zheng and Lin, Stephen and Guo, Baining},
  booktitle={ICCV},
  year={2021}
}