ZeroEC

Official implementation of ZeroEC: A Zero-Reference Framework for Robust Exposure Correction via Hallucination Consistency Learning.

Paper link will be added when available.

Ao Li¹, Zhenyu Wang^2,*, Mingtao Feng¹, Tao Huang², Yufan Zhu¹, Yuxin Feng², Fangfang Wu³, Weisheng Dong¹

¹School of Artificial Intelligence, Xidian University
²Hangzhou Institute of Technology, Xidian University
³School of Computer Science and Technology, Xidian University
^*Corresponding author

Overview

ZeroEC is a zero-reference learning framework for general exposure correction. It does not require curated paired or unpaired exposure-specific training data. Instead, it trains correction models from ordinary natural images by combining:

• A modified camera model that continuously samples exposure gain and camera response parameters to synthesize both over- and under-exposure degradations.
• An invert-clip-invert strategy for realistic under-exposure simulation and shadow detail loss.
• Hallucination consistency learning, where the corrected image is re-degraded with the opposite exposure and constrained against the deterministic opposite-exposure target.
• A model-agnostic pipeline instantiated with both a one-step diffusion backbone and a lightweight CoTF/LUT backbone.

Code Structure

• OSEDiff/: one-step diffusion instantiation of ZeroEC, following the OSEDiff-style backbone with LoRA training.
• CoTF/: CoTF/LUT instantiation of ZeroEC, implemented in BasicSR style with the same on-the-fly degradation and hallucination consistency objective.

Each subfolder has its own README.md and requirements.txt.

Datasets

ZeroEC trains on ordinary natural images, such as Flickr2K or MS COCO, rather than exposure-specific datasets. Evaluation in the paper uses:

• MSEC: https://github.com/mahmoudnafifi/Exposure_Correction
• SICEV2
• LOL v1

Datasets are not included in this repository. Download them separately and update the option files or command-line paths locally.

Installation

Use the environment for the branch you want to run.

OSEDiff Branch

cd OSEDiff
pip install -r requirements.txt

CoTF Branch

cd CoTF
pip install -r requirements.txt

Release Notes

This repository intentionally excludes datasets, pretrained weights, checkpoints, logs, cached files, and archived experiment outputs. Download datasets and pretrained models separately.