Research map / Image generation / Image restoration
Exposure-slot: Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware 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 Exposure-slot, a U-shaped encoder-decoder for image exposure correction that uses hierarchical slot-in-slot attention to partition exposure regions and learnable exposure-aware prompts to enhance features.
Problem
Image exposure correction (under- and over-exposure) where existing methods that separate features by physical properties do not yield optimal features, causing artifacts and poor exposure adjustment in regions like skies.
Approach
Extends Slot Attention to a hierarchical slot-in-slot structure (main-slots and sub-slots) whose attention maps act as exposure-aware region maps that weight learnable exposure-specific prompts, fed via cross-attention into the decoder; an auxiliary slot decoder supervises slots during training.
Key results
- Exposure-slot reaches 22.81 dB PSNR on SICE average, versus 20.41 dB for CSEC (SICE average gain reported as 1.85 dB), with 1.229M parameters.
- On LCDP, Exposure-slot achieves 23.18 dB PSNR average vs 22.73 dB for CSEC (about 0.4 dB gain).
- Color metrics improve to 6.94 ∆E2000 and 8.89 ∆Eab on SICE, versus 8.68 and 11.02 for the best prior method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Exposure-slot instead of FECNet | Exposure correction · SICE Avg. | PSNR: 20.96 -> 22.81 | Helped |
| use Exposure-slot instead of FECNet | Exposure correction · MSEC Avg. | PSNR: 23.12 -> 23.18 | Helped |
| use Exposure-slot instead of CSEC | Exposure correction · LCDP Avg. | PSNR: 23.63 -> 24.03 | Helped |
| add Decslot to the configuration without modules (ablation) | Exposure-slot · SICE | PSNR: 21.92 -> 22.02 | Helped |
| add prompts to the configuration without modules (ablation) | Exposure-slot · SICE | PSNR: 21.92 -> 22.09 | Helped |
| add 2-level SSAB alone to the configuration without modules (ablation) | Exposure-slot · SICE | PSNR: 21.92 -> 21.76 | Hurt |
| use the full method rather than the configuration without modules (ablation) | Exposure-slot · SICE | PSNR: 21.92 -> 22.81 | Helped |
| use 3-level rather than 2-level SSAB (ablation) | Exposure-slot · SICE | PSNR: 22.81 -> 23.06 | Helped |
| use 1-level rather than 2-level SSAB (ablation) | Exposure-slot · SICE | PSNR: 22.81 -> 22.02 | Hurt |
| set Kmain to 2 rather than 3 (ablation) | Exposure-slot · SICE | PSNR: 22.81 -> 22.61 | Hurt |
| set Kmain to 4 rather than 3 (ablation) | Exposure-slot · SICE | PSNR: 22.81 -> 22.36 | Hurt |
| set Ksub to 6 rather than 7 (ablation) | Exposure-slot · SICE | PSNR: 22.81 -> 21.92 | Hurt |
Surprising findings
- Adding a slot decoder and prompts alone (without SSAB) gives mixed results, with prompts lowering SSIM in one configuration.
Limitations
- Additional hierarchy levels increase time complexity exponentially.
- Gains on some under-exposed SSIM metrics are not best, so improvements are not uniform across all settings.
Experimental setup
Datasets: SICE MSEC LCDP
Benchmarks: SICE MSEC LCDP
Metrics: PSNR SSIM ∆E2000 ∆Eab parameter count runtime
Baselines: CSEC FECNet ECLNet DRBN-ENC MSEC LCDPNet PromptIR ERL MSEC+DA RetinexNet ZeroDCE RUAS
Builds on: Slot Attention PromptIR CoOp
Closest papers by meaning
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.87)
- Adaptive Slot Attention: Object Discovery with Dynamic Slot Number (CVPR 2024 · similarity 0.84)
- Identifiable Object-Centric Representation Learning via Probabilistic Slot Attention (NeurIPS 2024 · similarity 0.84)
- MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning (NeurIPS 2025 · similarity 0.84)
- GLASS: Guided Latent Slot Diffusion for Object-Centric Learning (CVPR 2025 · similarity 0.83)
- Grounded Object-Centric Learning (ICLR 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.