LibreHVI-CIDNet includes code adapted from HVI-CIDNet:

Source: https://github.com/Fediory/HVI-CIDNet
Commit: eb43d7d91e9a336c66856824ff9e4603ae41f408
License: MIT
Copyright (c) 2024 Fediory Feng

The inference port covers net/CIDNet.py, net/HVI_transform.py, net/LCA.py,
and net/transformer_utils.py. Module and parameter names mirror the upstream
network. The upstream MIT einops rearrangements are expressed with equivalent
native PyTorch reshape operations; training, losses, metrics, and dataset code
are not included.

Source path behaviorally referenced: eval_hf.py (RGB [0,1] preprocessing,
right/bottom reflection padding to a multiple of 8, gamma/control application,
output clamping, and crop to the original canvas). The family helper is a
native implementation rather than a copy of that file.

Architectural lineage disclosed by the upstream project:

- Restormer, https://github.com/swz30/Restormer, commit
  68dc6ac472db26f16361150cb7a96a1bc87da93f, MIT. Copyright (c) 2022 Syed
  Waqas Zamir and contributors.
- ConvNeXt, https://github.com/facebookresearch/ConvNeXt, commit
  048efcea897d999aed302f2639b6270aedf8d4c8, MIT. Copyright (c) Meta
  Platforms, Inc. and affiliates.
- einops, https://github.com/arogozhnikov/einops, commit
  e0d5eb4fd535945ff65d309206d0f1754a926821, MIT. Copyright (c) 2018 Alex
  Rogozhnikov. The port has no einops runtime dependency; native PyTorch
  implements the required reshapes.

Released checkpoint used by LibreYOLO:

- Repository: https://huggingface.co/Fediory/HVI-CIDNet-Generalization
- Revision: 51481ef2546f870060c43eb6d6525399f5b3d2d3
- File: model.safetensors (7,920,332 bytes)
- SHA-256: 2291407125e809cc9c0614cc2d010d21d309a66eb3da33e1ee2386a68fa05894
- Publisher-declared checkpoint license: MIT

Training-data caveat
--------------------

The upstream author states that the Generalization checkpoint was trained on
LOLv2-Synthetic with random gamma. The canonical LOLv2 source repository
(https://github.com/flyywh/CVPR-2020-Semi-Low-Light, commit
9f383decbd2717ab37bb9e4c133b3a0bf98ba638) publishes no dataset LICENSE or
explicit use terms. The checkpoint publisher directly grants MIT terms for the
released model artifact, but LibreYOLO has not independently established the
training dataset's commercial-use terms. No LOLv2 images or other training data
are included or redistributed by LibreYOLO.
