Research map / Image generation / Image restoration
Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images
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 DehazeXL, an end-to-end haze removal model that tokenizes fixed-size image patches (Swin Transformer V2 encoder/decoder) and fuses global context with an efficient LLM-style global attention bottleneck, plus a new 8KDehaze dataset and a Dehazing Attribution Map.
Problem
Haze removal on large, high-resolution images is hard: deep dehazing models run out of GPU memory, so they resort to slicing (losing global context) or downsampling (losing high-frequency detail). There is also a lack of ultra-high-resolution haze removal benchmarks.
Approach
Partition the hazy image into fixed-size patches encoded to tokens by a shared encoder, inject global context into tokens with a bottleneck global attention module (locality-sensitive hashing / Hyper Attention plus RMSNorm), and decode patches sequentially in mini-batches so memory is decoupled from image size.
Key results
- DehazeXL reaches 32.35 dB PSNR / 0.9863 SSIM on 8KDehaze, the best among evaluated methods (next best DehazeFormer-b 26.83 dB, ConvIR-b 26.93 dB).
- On 4KID DehazeXL achieves 26.62 dB PSNR vs 21.92 dB for the next best (ConvIR-b), and on O-HAZE 21.49 dB.
- DehazeXL can infer 10240x10240 images with about 21 GB of GPU memory using FP16.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DehazeXL instead of the strongest listed baseline | DehazeXL · 8KDehaze | PSNR: 26.93 -> 32.35 | Helped |
| Use DehazeXL instead of the strongest listed baseline | DehazeXL · 8KDehaze | SSIM: 0.9775 -> 0.9863 | Helped |
| Use DehazeXL instead of the strongest listed baseline | DehazeXL · 4KID | PSNR: 23.22 -> 26.62 | Helped |
| Use DehazeXL instead of the strongest listed baseline | DehazeXL · 4KID | SSIM: 0.9063 -> 0.9073 | Helped |
| Use DehazeXL instead of the strongest listed baseline | DehazeXL · O-HAZE | PSNR: 20.67 -> 21.49 | Helped |
| Use DehazeXL instead of the strongest listed baseline | DehazeXL · O-HAZE | SSIM: 0.7293 -> 0.7348 | Helped |
| Increase Bottleneck depth from 1 to 2 (ablation) | DehazeXL with Swin-T · 8KDehaze | PSNR: 31.61 -> 32.35 | Helped |
| Increase Bottleneck depth from 2 to 4 (ablation) | DehazeXL with Swin-T · 8KDehaze | PSNR: 32.35 -> 32.40 | Helped |
| Increase Bottleneck depth from 1 to 2 (ablation) | DehazeXL with Swin-S · 8KDehaze | PSNR: 31.89 -> 32.58 | Helped |
| Increase Bottleneck depth from 2 to 4 (ablation) | DehazeXL with Swin-S · 8KDehaze | PSNR: 32.58 -> 32.76 | Helped |
| Increase Bottleneck depth from 1 to 2 (ablation) | DehazeXL with Swin-B · 8KDehaze | PSNR: 32.10 -> 32.77 | Helped |
| Increase Bottleneck depth from 2 to 4 (ablation) | DehazeXL with Swin-B · 8KDehaze | PSNR: 32.77 -> 33.06 | Helped |
Surprising findings
- The 4KDehazing direct-inference variant performs much worse than its slicing variant on 8KDehaze (20.41 vs 25.81 dB PSNR).
Limitations
- Training and evaluation of the baselines were restricted to 2048x2048 crops or slicing, so comparisons on large images depend on baseline adaptations.
- Hazy images in 8KDehaze are synthetic, generated via atmospheric scattering model, so real-world transfer is only checked qualitatively on O-HAZE.
- Attribution analysis is qualitative (visual maps) without quantitative validation.
Experimental setup
Datasets: 8KDehaze 4KID O-HAZE
Benchmarks: 8KDehaze 4KID O-HAZE
Metrics: PSNR SSIM inference time
Baselines: 4KDehazing Dehamer C2PNet DehazeFormer MB-TaylorFormer ConvIR MixDehazeNet DEA-Net
Builds on: Swin Transformer V2 Hyper Attention Integrated Gradients Local Attribution Map
Closest papers by meaning
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.89)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.89)
- Real-world Image Dehazing with Coherence-based Pseudo Labeling and Cooperative Unfolding Network (NeurIPS 2024 · similarity 0.87)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.87)
- Guided Real Image Dehazing Using YCbCr Color Space (AAAI 2025 · similarity 0.86)
- Dehaze-RetinexGAN: Real-World Image Dehazing via Retinex-based Generative Adversarial Network (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.