LibreDDColort-restore

LibreDDColort-restore

DDColor automatic image colorization with the ConvNeXt-T encoder, converted for LibreYOLO's existing restore task. The network predicts Lab chroma at 512 square and reconstructs RGB on the source canvas using the original luminance plane.

Checkpoint license and training-data terms are separate. The publisher declares this exact artifact Apache-2.0. It was trained on ImageNet and has ImageNet-22K initialization lineage; ImageNet's access agreement limits dataset use to non-commercial research and education. No ImageNet data is included here. DDColor's Artistic checkpoint is intentionally excluded because it also uses undisclosed private data.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreDDColort-restore.pt")
result = model("black-and-white.jpg")
result.restored.save("colorized.png")

Provenance

• Source repository: piddnad/ddcolor_paper_tiny
• Revision: cf9fd99c1d7472689ec7413441c1b799a51866a3
• Source file: pytorch_model.bin, 220,372,845 bytes
• Source SHA-256: 8a1277bc90a1bfbb6d2d83933a9a6bc821931879ca93e26e4fcec12165d41fce
• Converted SHA-256: 044254d616df5cc6935669ea05cf641dbd9af5b53a04096449b820d4e64c4a3d
• Architecture source: piddnad/DDColor at 2adb63f2656ac41cbdf7b894cddd94121a3faf13

Learned tensors are unchanged. Conversion adds LibreYOLO v1 checkpoint metadata. Network parity is exact (max_abs_diff=0), and the complete OpenCV Lab pipeline is pixel-identical to the pinned reference.

License

The exact source artifact is publisher-declared Apache-2.0. See LICENSE and NOTICE. The ImageNet data caveat above is retained as provenance and is not erased by conversion.