Research map / Image generation / Image restoration
MESS: Multi-Exposure Sequence Synthesis for Generalizable Image Enhancement
Ruodai Cui, Shuaizheng Liu, Rongyuan Wu, Lei Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MESS, a method for synthesizing multi-exposure image sequences to create large-scale, diverse training data for generalizable image enhancement under adverse illumination.
Problem
Existing image enhancement models (low-light enhancement, exposure correction, backlit enhancement) generalize poorly across scenarios, largely because training requires expensive, manually captured, precisely aligned image pairs that are scarce and limited in diversity.
Approach
Synthesizes multi-exposure image sequences to supply large-scale, diverse training data for enhancement tasks instead of relying on manually captured aligned pairs. Full method details are not available from the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on MESS-generated data instead of in-domain real datasets | FourLLIE · 5 enhancement test sets | PSNR (average): 19.965 -> 21.923 (vs SICE) | Improved |
| train on MESS-generated data instead of RAW-synthesized MSEC data | FourLLIE · 5 enhancement test sets | PSNR (average): 20.079 -> 21.923 | Improved |
| train on MESS data instead of RGB-synthesized UEC-based data | FourLLIE · 5 enhancement test sets | PSNR (average): 19.645 -> 21.923 | Improved |
| train on MESS data instead of in-domain data on its own test set | FourLLIE · UHDLL, SICE | PSNR: 23.694 -> 23.416 (UHDLL); 22.385 -> 21.769 (SICE) | Hurt |
| scale MESS training set from 5K to 500K pairs (ablation) | FourLLIE · 5 test sets average | PSNR: 20.220 -> 21.923 | Improved |
| train on 5K MESS pairs instead of real LSRW/SICE of similar size (ablation) | FourLLIE · 5 test sets average | PSNR: 19.965 -> 20.220 | Improved |
| replace RAW-only training data with MESS-RGB data of matched size (ablation) | FourLLIE · 5 test sets average | PSNR: 20.764 -> 21.083 | Improved |
| remove frequency-based enhancement, diffusion detail, exposure conditioning or representation consistency loss (ablation) | MESS generator · downstream FourLLIE, 5 test sets average | FR/NR metrics | Improved |
| use DWT instead of FFT for frequency modeling (ablation) | MESS generator · downstream FourLLIE | FR/NR metrics | Improved |
| inject exposure via DWT-FiLM and content descriptors instead of plain concatenation or no condition (ablation) | MESS renderer · downstream FourLLIE | FR/NR metrics | Improved |
Closest papers by meaning
- Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling (CVPR 2024 · similarity 0.87)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.87)
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.86)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.86)
- ECMamba: Consolidating Selective State Space Model with Retinex Guidance for Efficient Multiple Exposure Correction (NeurIPS 2024 · similarity 0.86)
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
