RGB stumbles in the dark; HVI was built for it

What it does

HVI-CIDNet is a transformer-based low-light image enhancer that converts images into a custom HVI color space before processing. It uses a Lighten Cross-Attention (LCA) block to recover detail from extremely dark scenes, and the authors provide pre-trained weights and reproducible benchmarks across a half-dozen datasets including LOLv1/v2, Sony Total Dark, and LOL-Blur. A Gradio demo on Hugging Face lets you upload a photo and judge the results without writing any code.

The interesting bit

The core pitch is that a dedicated color space matters for extreme low-light restoration, so the authors designed HVI specifically for the task rather than relying on off-the-shelf representations. They also use random gamma curves during training to improve cross-dataset generalization, which is a practical trick tucked into train.py.

Key highlights

• CVPR 2025 official implementation with weights hosted on Hugging Face.
• Evaluated on paired datasets (LOLv1, LOLv2, Sony, SICE, FiveK) and unpaired ones (DICM, LIME, MEF, NPE, VV).
• Includes a Gradio web UI for quick visual tests.
• The authors’ NTIRE 2025 winning entry was a fusion network built on top of this base model.
• An upgraded “HVI-CIDNet+” variant for extreme darkness exists on ArXiv, though the code lives in a separate repo.

Caveats

• The README relies on images to explain the HVI color space and architecture; the actual text describing how HVI differs mathematically from conventional spaces is sparse.
• Several weight files for LOLv2_real are listed as “(lost)” in the results table.
• The follow-up HVI-CIDNet+ paper and code have moved to a different repository.

Verdict

Worth a look if you research low-light restoration or need a strong baseline with published weights. Skip it if you want a fully documented color-space specification or a drop-in production API; this is research code with a demo attached.