Research map / Image generation / Image restoration
Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range 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 a zero-shot, structure-preserving latent diffusion model for HDR tone mapping that trains only on LDR images, using MSCN structure as shared HDR/LDR knowledge plus a structure refinement step during sampling.
Problem
Tone mapping HDR images to high-quality LDR images lacks paired HDR/LDR training data, and existing adversarial methods trade off tone naturalness against structure preservation.
Approach
Decompose images into tone (local mean/std) and structure (MSCN coefficients), which are distributed nearly identically across HDR and LDR domains; train a conditional diffusion model on LDR images conditioned on MSCN and blurred luma, then at inference feed HDR MSCN maps and inject the original structure into intermediate denoising results (structure refinement operation).
Key results
- On HDRPS (105 HDR images), the proposed method reaches TMQI 0.8915 and NIQE 2.340, the best on both metrics versus Vinker et al. (0.8861, 3.174) and DeepTMO (0.88, 2.519).
- Ablation on SRO: default strategy (t0=10, T=20) gives TMQI 0.8915 versus 0.7210 with no SRO and 0.8260 with SRO at all iterations.
- JS divergence of MSCN coefficient distributions between HDR and LDR is 0.002, versus 0.365 for local means and 0.184 for local std.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed zero-shot structure-preserving diffusion method instead of Vinker et al. | HDR tone mapping · HDRPS | TMQI: 0.8861 -> 0.8915 | Helped |
| use the proposed zero-shot structure-preserving diffusion method instead of Vinker et al. | HDR tone mapping · HDRPS | NIQE: 3.174 -> 2.340 | Helped |
| add the structure-refinement operation during sampling (ablation) | Proposed diffusion method · HDRPS | TMQI: 0.7210 -> 0.8915 | Helped |
| apply SRO at the last iteration rather than using the default strategy (ablation) | Proposed diffusion method · HDRPS | TMQI: 0.8915 -> 0.8257 | Hurt |
| apply SRO at earlier iterations rather than using the default strategy (ablation) | Proposed diffusion method · HDRPS | TMQI: 0.8915 -> 0.7297 | Hurt |
| apply SRO at all iterations rather than using the default strategy (ablation) | Proposed diffusion method · HDRPS | TMQI: 0.8915 -> 0.8260 | Hurt |
| add a blurred luma map alongside the MSCN map (ablation) | Dual-control network · HDR tone mapping | luminance realism | Helped |
| input the low-pass luma map at a coarser scale instead of the same scale (ablation) | Dual-control network · HDR tone mapping | image detail | Helped |
Surprising findings
- Applying structure refinement at every sampling iteration reduces hallucinated artifacts but also reduces contrast and edge intensity, so it is disabled in the final timesteps.
Limitations
- Evaluation is limited to one main benchmark (HDRPS) with only TMQI and NIQE metrics and no human study.
- Generalization to other tasks (VIS-NIR fusion) is only shown qualitatively.
Experimental setup
Datasets: Flickr2K HDRPS HDR+ VIS-NIR Scene
Benchmarks: HDRPS
Metrics: TMQI NIQE
Baselines: DeepTMO Vinker et al. (2021 tone mapping) LA-Net Liang et al. Shibata et al. ETMO
Builds on: Stable Diffusion DDIM DDPM MSCN normalization
Closest papers by meaning
- Generating Content for HDR Deghosting from Frequency View (CVPR 2024 · similarity 0.90)
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.90)
- Learning Differential Pyramid Representation for Tone Mapping (NeurIPS 2025 · similarity 0.88)
- Real-Time Scene-Adaptive Tone Mapping for High-Dynamic Range Object Detection (NeurIPS 2025 · similarity 0.88)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.87)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.87)
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.