What it does

HiDT takes a photo and translates its daytime lighting condition—think morning to sunset, overcast to clear—while preserving the underlying scene. It runs at high resolution and includes an optional enhancer network to clean up artifacts. The repo ships with pretrained models for daytime translation and a WikiArt style transfer variant, plus a one-click Colab notebook for the inference-shy.

The interesting bit

The unsupervised angle is the real trick: the model learns to disentangle “content” from “lighting/style” without anyone labeling which photos are sunrise versus noon. The paper (CVPR 2020 Oral) proposes a hierarchical design that handles 1024×1024 images without collapsing into mud—still a genuine pain point in GAN-based translation.

Key highlights

• Inference-only release; training code is not included
• Two enhancement modes: a dedicated upsampling generator (G_enh) or fully-convolutional raw output
• Pretrained checkpoints for daytime and WikiArt domains
• One Colab notebook for browser-based experimentation
• Python ≥3.7, PyTorch, tested with conda

Caveats

• “Only inference part” means you cannot retrain on your own domains without reimplementing the training loop
• The README contains a typo (“requirenments”) and sparse troubleshooting guidance
• No explicit hardware requirements listed, though 1024×1024 GAN inference implies a decent GPU

Verdict

Grab it if you need to relight outdoor photography or prototype style transfer without curating datasets. Skip it if you need fine-grained control over training or are working on non-visual domains—this is a narrow, well-executed tool, not a framework.