One-stage Retinex transformer handles 4000×6000 low-light frames

What it does Retinexformer is a one-stage transformer that performs low-light image enhancement using Retinex theory. The repository packages the ICCV 2023 method as a full toolbox, supporting training and inference on over 15 standard benchmarks and handling images up to 4000×6000 pixels via an adaptive split-and-test strategy. It also supplies pre-trained weights, training logs, and comparison results against other baselines.

The interesting bit Instead of stacking separate illumination-estimation and noise-removal stages, the network does everything in one pass. That simplicity seems to scale well: follow-up works built on this framework have taken first place in the NTIRE 2025 and 2026 low-light challenges, while the original method placed second in NTIRE 2024.

Key highlights

• Supports 15+ benchmarks including LOL-v1/v2, SID, SMID, SDSD, MIT Adobe FiveK, and NTIRE 2024 data
• Handles extremely high-resolution enhancement (up to 4000×6000) with an adaptive split-and-test mode
• Includes distributed data parallel and mixed-precision training for larger models
• Provides results and models for ExDark nighttime object detection downstream tasks
• ICCV 2023 top-10 cited paper; NTIRE 2024 runner-up with later derivatives winning NTIRE 2025 and 2026

Caveats

• The authors explicitly discourage one of the provided test settings (used by KinD, LLFlow, and diffusion-model papers) because it averages the ground-truth mean to inflate scores

Verdict Computer-vision researchers working on low-light restoration or nighttime detection should keep this bookmarked; casual users looking for a drop-in phone app will find it stays firmly in the PyTorch/conda realm.