LibreLaMab-restore

Mask-guided LaMa inpainting under LibreYOLO's existing restore task. This checkpoint embeds the exact fixed-512 QDQ ONNX graph published by OpenCV Zoo. Every nonzero mask pixel means fill; zero means preserve.

OpenCV Zoo publishes the exact ONNX artifact under Apache-2.0. The graph was trained on Places365-Challenge, whose image-download terms are limited to non-commercial research and education. Those terms do not expressly define trained-model redistribution, and commercial training-data clearance was not independently established. No Places365 image is included here.

pip install "libreyolo[onnx]"

from libreyolo import LibreYOLO

model = LibreYOLO("LibreLaMab-restore.pt")
result = model("photo.jpg", mask="erase-mask.png")
result.restored.save("inpainted.png")

The image and mask must share one canvas. LibreYOLO restores the source size and copies every unmasked source pixel back exactly.

Provenance

• Artifact repository: opencv/inpainting_lama
• Revision: aee6d22f0a13e5e35af1c9a1c3afd62841fc6f3f
• Source file: inpainting_lama_2025jan.onnx, 92,591,623 bytes
• Source SHA-256: 7df918ac3921d3daf0aae1d219776cf0dc4e4935f035af81841b40adcf74fdf2
• Converted SHA-256: 2cb9a47f76ced186749b641a0ba5ff86e534203ad34c362d30abfad5b9e3a8be
• OpenCV Zoo source: opencv/opencv_zoo at 966be4cb6eefc2a4f586ed0581bf4bbaa2533070

The ONNX bytes are unchanged and are stored as one persistent uint8 state-dict buffer. LibreYOLO verifies the embedded bytes before creating the ONNX Runtime session. OpenCV 4.x cannot execute this opset-21 graph; the LibreYOLO runtime requires ONNX Runtime 1.18 or newer.

License

OpenCV Zoo declares all files in the model directory Apache-2.0. See LICENSE and NOTICE. The Places365 caveat above is retained separately.