Research map / Image generation / Image restoration
CLIP-RestoreX: Restore Image Structure and Perception in Exposure Correction
Read the paperCheck your idea against 108,967 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 CLIP-RestoreX, which uses CLIP image-encoder features from shallow and deep blocks as structural and perceptual priors, enhanced by a frequency-domain diffusion model (FFEDM) and injected into a Restormer for exposure correction.
Problem
Exposure correction of severely under- and over-exposed images, where existing methods fail to restore damaged structural and perceptual information simultaneously.
Approach
Align CLIP-based structural (shallow blocks) and perceptual (deep blocks) features of the degraded input with those of the ground-truth, using a frequency-domain conditional diffusion model to enhance the degraded priors and injecting them into Restormer blocks, plus a CLIP-feature-based SAP loss.
Key results
- On LCDP, the method reaches 24.95 PSNR / 0.8864 SSIM, versus 24.27 / 0.8621 for the best prior method (RECNet).
- On SICE, the method reaches 23.21 PSNR / 0.7302 SSIM, versus 22.11 / 0.7102 for the best prior method (RECNet).
- Ablation on LCDP: adding CLIP priors raises PSNR from 23.67 to 24.14 with FFEDM-free diffusion; full model with FFEDM and SAP loss reaches 24.95.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CLIP-RestoreX instead of LACT | Exposure correction · LCDP | PSNR: 24.27 -> 24.95 | Helped |
| use CLIP-RestoreX instead of LACT | Exposure correction · LCDP | SSIM: 0.8621 -> 0.8864 | Helped |
| use CLIP-RestoreX instead of LACT | Exposure correction · MSEC | PSNR: 23.53 -> 23.57 | Helped |
| use CLIP-RestoreX instead of CoTF | Exposure correction · MSEC | SSIM: 0.8733 -> 0.8764 | Helped |
| use CLIP-RestoreX instead of LACT | Exposure correction · SICE | PSNR: 22.11 -> 23.21 | Helped |
| use CLIP-RestoreX instead of LACT | Exposure correction · SICE | SSIM: 0.7102 -> 0.7302 | Helped |
| use shallow or deep CLIP blocks alone instead of combining shallow and deep blocks (ablation) | CLIP-RestoreX on LCDP · LCDP | PSNR: 24.75 -> 24.42; 24.75 -> 24.63 | Hurt |
| use shallow or deep CLIP blocks alone instead of combining shallow and deep blocks (ablation) | CLIP-RestoreX on LCDP · LCDP | SSIM: 0.8835 -> 0.8824; 0.8835 -> 0.8815 | Hurt |
| add SAP loss to the model without CLIP priors, diffusion model, or FFEDM (ablation) | CLIP-RestoreX on LCDP · LCDP | PSNR: 23.67 -> 24.10 | Helped |
| add CLIP priors to the diffusion model (ablation) | CLIP-RestoreX on LCDP · LCDP | PSNR: 24.14 -> 24.75 | Helped |
| replace the diffusion model with FFEDM (ablation) | CLIP-RestoreX on LCDP · LCDP | PSNR: 24.75 -> 24.86 | Helped |
| add SAP loss to CLIP priors and FFEDM (ablation) | CLIP-RestoreX on LCDP · LCDP | PSNR: 24.86 -> 24.95 | Helped |
Surprising findings
- Shallow CLIP blocks (1-4) alone or deep blocks (9-12) alone underperform the combined 3-6-9-12 setting, while shallow blocks of the ImageNet-pretrained ViT-B-32 do not show the same brightness-robust behavior as CLIP.
Limitations
- Evaluated only with PSNR and SSIM on three datasets; no perceptual metrics reported.
- Two-stage training with different crop sizes adds training complexity.
Experimental setup
Datasets: LCDP MSEC SICE
Benchmarks: LCDP MSEC SICE
Metrics: PSNR SSIM
Baselines: MSEC ENC-SID ENC-DRBN LCDPNet FECNet ECLNet LACT RECNet CoTF
Builds on: CLIP Restormer conditional denoising diffusion probabilistic model FECNet
Closest papers by meaning
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.85)
- InstructRestore: Region-Customized Image Restoration with Human Instructions (NeurIPS 2025 · similarity 0.83)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.83)
- Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model (ICLR 2025 · similarity 0.83)
- Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling (CVPR 2024 · similarity 0.83)
- Boosting Image Restoration via Priors from Pre-trained Models (CVPR 2024 · similarity 0.83)
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.