LibreHVICIDNett-restore

The 1.98M-parameter HVI-CIDNet Generalization checkpoint for adjustable low-light enhancement under LibreYOLO's existing restore task. Prediction preserves the source canvas and pads internally to a multiple of eight.

The publisher declares this exact checkpoint MIT. It was trained on LOLv2-Synthetic; the canonical LOLv2 source repository publishes no dataset license or explicit use terms. No training images are included here.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreHVICIDNett-restore.pt")
result = model(
    "night.jpg",
    gamma=1.0,
    saturation=1.0,
    intensity=1.0,
)
result.restored.save("enhanced.png")

All three controls must be positive. 1.0 reproduces the published Generalization checkpoint's evaluation configuration.

Provenance

• Source repository: Fediory/HVI-CIDNet-Generalization
• Revision: 51481ef2546f870060c43eb6d6525399f5b3d2d3
• Source file: model.safetensors, 7,920,332 bytes
• Source SHA-256: 2291407125e809cc9c0614cc2d010d21d309a66eb3da33e1ee2386a68fa05894
• Converted SHA-256: 145ced1ff039ed6c18ff6052dcf9a056e3fbf6b5594238facbbdf67045a855de
• Architecture source: Fediory/HVI-CIDNet at eb43d7d91e9a336c66856824ff9e4603ae41f408

Learned tensors are unchanged. Conversion adds LibreYOLO v1 checkpoint metadata. The native graph matches the pinned reference exactly (max_abs_diff=0).

License

The exact source artifact is publisher-declared MIT. See LICENSE and NOTICE. The unresolved LOLv2 dataset terms above remain visible as a separate provenance caveat.