UNICE: Training A Universal Image Contrast Enhancer

Ruodai Cui, Lei ZhangarXiv 2025 · DOI · arXiv

Abstract

UNICE (Universal Image Contrast Enhancer) addresses the fragmentation in image enhancement tasks by providing a single, robust model for low-light, over-exposure, and backlit corrections. By utilizing a massive dataset of RAW images to generate multi-exposure sequences (MES) and a one-step diffusion-based pipeline, UNICE achieves superior generalization without human-labeled ground truth. Experimental results demonstrate that UNICE often outperforms specialized state-of-the-art models and produces results preferred by humans over manual retouches.

Introduction

In computer vision, image enhancement is traditionally divided into specialized sub-tasks such as Low-Light Image Enhancement (LLIE), Exposure Correction (EC), and Backlit Image Enhancement (BIE). Specialized models for these tasks are often "brittle," failing when applied to datasets or lighting conditions outside their narrow training scope. Furthermore, training these models typically requires expensive and subjective human-edited "Ground Truth" (GT) images, which can lead to inconsistent results.

UNICE proposes a unified framework based on the physics of light. Rather than treating different lighting failures as separate problems, UNICE frames them as a single challenge of exposure and contrast adjustment, best handled by recovering the full dynamic range of a scene.

Methods

Automated Data Generation

To avoid the subjectivity of human retouching, the researchers utilized 48,361 public RAW images. RAW files contain significantly more light information than standard JPEGs. Using an emulated Image Signal Processor (ISP), each RAW image was rendered into a Multi-Exposure Sequence (MES)—a bracketed set of photos ranging from very dark to very bright. These were then blended using Multi-Exposure Fusion (MEF) to create high-quality "pseudo-ground truth" images that preserve details in both shadows and highlights.

The UNICE Pipeline

The model architecture consists of two primary stages built on the SD-Turbo backbone (a high-speed, one-step diffusion model) and fine-tuned using LoRA (Low-Rank Adaptation):

  1. MES-Net: This network takes a single standard image and predicts its corresponding multi-exposure sequence, effectively recovering missing dynamic range.
  2. MEF-Net: This network takes the predicted sequence and fuses it into a final, enhanced output.
Overview of the UNICE pipeline, detailing dataset generation from RAW images, the two-stage network training (MES-Net and MEF-Net), and the diffusion-based backbone.

Key Results

Superior Generalization

UNICE demonstrates remarkable universality. In cross-task evaluations (e.g., testing an exposure correction model on low-light data), UNICE significantly outperformed specialized state-of-the-art (SOTA) models. While specialized models saw performance drops when faced with unfamiliar tasks, UNICE remained steady across LLIE, EC, and backlit enhancement.

Beating the "Ground Truth"

In several no-reference quality metrics, UNICE’s outputs scored higher than the manually created ground truths in the datasets. This suggests that the automated fusion of RAW data provides more consistent and realistic results than manual human editing, which often introduces artifacts or unrealistic color casts.

Qualitative results showing UNICE's performance across diverse tasks including LLIE, Exposure Correction, and Backlit enhancement compared to ground truth.

Human Preference

A double-blind user study confirmed these findings, with volunteers preferring UNICE’s results over specialized competitors by a wide margin—receiving up to 83% of the votes in High Dynamic Range (L2HT) tasks.

Conclusion

UNICE represents a significant shift toward universal image enhancement. By focusing on recovering underlying light signals rather than mimicking human edits, it achieves a level of robustness and quality that specialized models lack. While its 965M parameter diffusion backbone presents challenges for mobile deployment, its ability to generalize across tasks without human supervision sets a new benchmark for the field. Future work will explore end-to-end joint training and extending this universal logic to video processing.

Summary updated

BibTeX
@misc{cui2025unice,
  author = {Ruodai Cui and Lei Zhang},
  title = {{UNICE: Training A Universal Image Contrast Enhancer}},
  year = {2025},
  eprint = {2507.17157},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.17157},
}
Figure