Consist-Retinex: One-Step Low-Light Image Enhancement via Consistency Training
Researchers propose Consist-Retinex, a method for low-light image enhancement that uses consistency models to achieve one-step restoration. The approach first employs a Retinex Transformer Decomposition Network to separate reflectance and illumination, then trains two conditional consistency models with a dual objective combining trajectory consistency and ground-truth component supervision. Adaptive noise-emphasized fixed-point sampling addresses instability in one-step inference. The method is published on arXiv (2512.08982) and aims to accelerate high-quality Retinex enhancement for deployment under strict latency budgets.
Key facts
- Consist-Retinex is a one-step low-light image enhancement method.
- It uses a Retinex Transformer Decomposition Network (TDN).
- Two conditional consistency models are trained with a dual objective.
- The dual objective combines trajectory consistency and paired ground-truth component supervision.
- Adaptive noise-emphasized fixed-point sampling is used.
- The method addresses instability in one-step inference.
- Published on arXiv with identifier 2512.08982.
- Aims to accelerate high-quality Retinex enhancement under strict latency budgets.
Entities
Institutions
- arXiv