Research map / Image generation / Image restoration
Learning Differential Pyramid Representation for Tone Mapping
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 DPRNet, an end-to-end tone mapping network with a learnable differential pyramid, global/local tone modules on downsampled inputs, and iterative coarse-to-fine detail enhancement, plus a new HDRI Haven dataset.
Problem
Tone mapping of HDR images loses high-frequency detail when operating on downsampled inputs and fixed handcrafted pyramids, and global and local tone consistency are modeled disjointly, causing flat results and halo artifacts.
Approach
A learnable differential pyramid generalizes Laplacian/DoG pyramids with content-adaptive differencing; it is combined with a global tone perception module (modulated conv), a local tone tuning module (patch-wise basis 3D LUTs), and an iterative coarse-to-fine detail enhancement module.
Key results
- On HDR+ at original 4K, DPRNet reaches 27.71 dB PSNR, 2.39 dB above LLFLUT (25.32 dB), with 212K parameters.
- On HDRI Haven at 4K, DPRNet reaches 28.30 dB PSNR, over 3 dB above the next best method (RECNet 25.29 dB per table).
- In a user study with 25 participants, DPRNet results were ranked first or second in 39.56% of cases, close to ground truth at 39.64%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LTT (ablation) | DPRNet · HDR+ (480p) | PSNR: 26.36 -> 27.95 | Helped |
| add GTP (ablation) | DPRNet · HDR+ (480p) | PSNR: 26.17 -> 27.95 | Helped |
| add LDP (ablation) | DPRNet · HDR+ (480p) | PSNR: 25.90 -> 27.95 | Helped |
| add IDE (ablation) | DPRNet · HDR+ (480p) | PSNR: 25.35 -> 27.95 | Helped |
| add SSIM loss (ablation) | DPRNet training objective · HDR+ (480p) | PSNR: 27.85 -> 27.95 | Helped |
| add high-frequency loss (ablation) | DPRNet training objective · HDR+ (480p) | PSNR: 27.74 -> 27.95 | Helped |
| add perceptual loss (ablation) | DPRNet training objective · HDR+ (480p) | PSNR: 27.76 -> 27.95 | Helped |
| use n=4 rather than n=3 pyramid levels (ablation) | DPRNet · HDR+ (480p) | PSNR: 27.11 -> 27.95 | Helped |
| use n=4 rather than n=5 pyramid levels (ablation) | DPRNet · HDR+ (480p) | PSNR: 27.02 -> 27.95 | Helped |
| set LUT to m=6 and S_t=9 rather than m=8 and S_t=9 (ablation) | DPRNet LTT · HDR+ (480p) | PSNR: 27.92 -> 27.95 | Helped |
| set LUT to m=6 and S_t=9 rather than m=6 and S_t=17 (ablation) | DPRNet LTT · HDR+ (480p) | PSNR: 27.98 -> 27.95 | No clear effect |
| use LTT grid size N=4 rather than N=2 (ablation) | DPRNet LTT · HDR+ (480p) | PSNR: 25.77 -> 27.95 | Helped |
Surprising findings
- Training on HDRI Haven instead of HDR+ improves TMQI on real-world HDR Survey (0.9245 vs 0.8157) and UVTM (0.9540 vs 0.8476) test sets.
- The method underperforms state-of-the-art on the MIT FiveK dataset, attributed to DSLR noise and inconsistent reference retouching.
Limitations
- Performance drops on MIT FiveK due to sensor noise and inconsistent reference images.
- Pyramid depth beyond 4 levels degrades performance, indicating difficulty reconstructing high frequencies at deep scales.
- Diffusion-based alternatives are hard to deploy on mobile, while the proposed method's real-time/mobile deployment is not evaluated.
Experimental setup
Datasets: HDR+ Burst Photography HDRI Haven
Benchmarks: HDR+ (480p and 4K) HDRI Haven (480p and 4K) HDR Survey UVTM
Metrics: PSNR SSIM TMQI LPIPS deltaE
Baselines: UPE HDRNet CSRNet DeepLPF 3D LUT LPTN CLUT sLUT SepLUT LLFLUT CODEFormer CoTF
Builds on: Laplacian pyramid Difference-of-Gaussian 3D LUT AdaInt AdamW
Closest papers by meaning
- Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping (CVPR 2024 · similarity 0.88)
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.85)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.84)
- Real-Time Scene-Adaptive Tone Mapping for High-Dynamic Range Object Detection (NeurIPS 2025 · similarity 0.83)
- Plug-and-Play Tri-Branch Invertible Block for Image Rescaling (AAAI 2025 · similarity 0.83)
- Learning with Unreliability: Fast Few-shot Voxel Radiance Fields with Relative Geometric Consistency (CVPR 2024 · similarity 0.82)
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.