LibreViTMattes-matte

Pretrained checkpoint restriction

NON-COMMERCIAL. This checkpoint was trained on Adobe Composition-1k. Adobe's Deep Image Matting Dataset License Agreement says models may only be used and distributed for non-commercial purposes and may not be sold. Permitted distribution must retain the required Deep Image Matting CVPR 2017 attribution. Read LICENSE before use or redistribution.

The restriction applies to this pretrained checkpoint, not to LibreYOLO's MIT code or to independently trained ViTMatte weights with suitable provenance.

Use

LibreViTMatte provides trimap-guided alpha matting under LibreYOLO's existing matte task. The trimap must use exactly 0/128/255 (or normalized 0/0.5/1) for known background, unknown, and known foreground.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreViTMattes-matte.pt")
result = model("portrait.jpg", trimap="trimap.png")
alpha = result.matte.array

The result is a soft float32 alpha matte on the source canvas. Known background and foreground pixels are forced to exactly zero and one.

Provenance

• Source repository: hustvl/vitmatte-small-composition-1k
• Revision: 6a58ad7646403c1df626fbd746900aec7361ea1d
• Source file: model.safetensors, 103,294,572 bytes
• Source SHA-256: bda9289db1bb6762d978b42d1c62ae3f34daf7497171a347a1d09657efd788cb
• Converted SHA-256: 20e5e41b601380d65ab7d685520ceba7b94456d097f1fbf2f3b8dc87ab0e27c6
• Inference port source: huggingface/transformers at 7d6354e04794f3246bf9a0faf4fead080edeebb6 (Apache-2.0)
• Architecture lineage: hustvl/ViTMatte at f48208ceb0d172ab0c1c528bba33cc1159596407 (MIT)

Learned tensors are unchanged. Conversion adds LibreYOLO v1 checkpoint metadata. Preprocessing, raw alpha, and constrained-alpha parity with the pinned Transformers reference are bit-exact.

Required attribution

Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep Image Matting. Proceedings of the Conference on Computer Vision and Pattern Recognition, 2017.

See NOTICE for complete model and code provenance.