Research map / Image generation / Image restoration
SelfPromer: Self-Prompt Dehazing Transformers with Depth-Consistency
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 SelfPromer, a VQGAN-based dehazing Transformer whose prompt is built from depth differences between hazy and clear feature maps, plus a continuous self-prompt inference that iteratively refines dehazed outputs.
Problem
Image dehazing leaves haze residuals that degrade perceptual quality; at inference clear images are unavailable, so depth-consistency guidance cannot directly use ground truth.
Approach
Build a text-free prompt from feature-level depth differences (monocular depth of hazy input vs. clear image, encoded by frozen VQGAN encoder), inject it via prompt embedding (additive to features) and prompt attention (additive to queries), and at test time replace the clear image with the model's own previous dehazed output in an iterative self-prompt loop.
Key results
- On SOTS-indoor, the 3-step variant reaches NIQE 4.054, PI 3.857, PIQE 27.927, best among compared methods (Dehamer PIQE 32.446 etc.).
- On SOTS-outdoor, the 1-step variant achieves NIQE 2.646 and PI 2.003, best on both perception metrics.
- On the real-world dataset, the 3-step variant gets NIQE 4.062, PI 3.391, PIQE 14.026, best among compared methods; PSNR/SSIM decrease with more iterations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 3rd prompt inference instead of 1st prompt inference | SelfPromer · SOTS-indoor | NIQE: 4.252 -> 4.054 | Helped |
| use 3rd prompt inference instead of 1st prompt inference | SelfPromer · SOTS-outdoor | NIQE: 2.646 -> 2.685 | Hurt |
| use 3rd prompt inference instead of 1st prompt inference | SelfPromer · SOTS-outdoor | PIQE: 6.577 -> 6.151 | Helped |
| use feature-level rather than image-level depth difference as the prompt (ablation) | SelfPromer · SOTS-indoor | NIQE: 4.901 -> 4.252 | Helped |
| use feature-level depth-difference prompting instead of concatenating image and depth features (ablation) | SelfPromer · SOTS-indoor | NIQE: 4.362 -> 4.252 | Helped |
| add the depth-difference prompt (ablation) | SelfPromer · SOTS-indoor | NIQE: 4.258 -> 4.252 | Helped |
| add prompt embedding instead of omitting embedding (ablation) | SelfPromer · SOTS-indoor | NIQE: 4.410 -> 4.252 | Helped |
| use prompt embedding instead of position embedding (ablation) | SelfPromer · SOTS-indoor | NIQE: 4.267 -> 4.252 | Helped |
| use prompt attention instead of regular attention (ablation) | SelfPromer · SOTS-indoor | NIQE: 4.300 -> 4.252 | Helped |
| increase continuous self-prompt inference to 3 steps | SelfPromer · SOTS-indoor | NIQE | Helped |
| use continuous self-prompt inference instead of recurrent dehazing | SelfPromer · SOTS-indoor | NIQE | Helped |
| use the 2nd or 3rd self-prompt instead of GT guidance | SelfPromer · SOTS-indoor | NIQE | Helped |
Surprising findings
- Distortion metrics (PSNR/SSIM) are not aligned with perceptual quality; methods with high PSNR such as Dehamer and D4 produce less natural results.
- Iterative self-prompt steps reduce PSNR and SSIM even as perception metrics improve, and GT-guided prompting underperforms the 2nd and 3rd self-prompt steps.
Limitations
- Performance depends on depth estimates; continuous self-prompt may fail when depth difference between iterations is not significant.
- Slight degradation in NIQE and PI on SOTS-outdoor with multiple prompt steps.
- Multi-iteration inference is slower than single-pass feed-forward models.
Experimental setup
Datasets: RESIDE ITS UnannotatedHazyImages
Benchmarks: SOTS-indoor SOTS-outdoor UnannotatedHazyImages
Metrics: NIQE PI PIQE PSNR SSIM
Baselines: GridNet PFDN UHD PSD Uformer Restormer D4 Dehamer CodeFormer
Builds on: VQGAN CodeFormer FeMaSR Uformer Restormer
Closest papers by meaning
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.90)
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.87)
- Unsupervised Blind Image Deblurring Based on Self-Enhancement (CVPR 2024 · similarity 0.87)
- Depth Information Assisted Collaborative Mutual Promotion Network for Single Image Dehazing (CVPR 2024 · similarity 0.87)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (AAAI 2025 · similarity 0.86)
- NightHaze: Nighttime Image Dehazing via Self-Prior Learning (AAAI 2025 · 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.