Research map / Image generation / Image restoration
Perceptual Assessment and Optimization of HDR Image Rendering
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 training-free HDR image quality metrics that decompose an HDR image into a multi-exposure LDR stack via an inverse display model and apply existing LDR metrics, then uses the metric as a loss for HDR novel view synthesis.
Problem
Assessing the perceptual quality of HDR image rendering is underexplored; existing quality models are designed for LDR images and align poorly with human HDR perception, and HDR-VDP models are complex and non-differentiable.
Approach
Use a gain-offset-gamma inverse display model to map an HDR image into a stack of LDR images at different exposures, then score each with well-established LDR metrics, weighting by well-exposedness and optionally optimizing exposure values to compensate luminance shifts, with no HDR human-rating recalibration.
Key results
- The Q-DISTS instantiation achieves the best SRCC/PLCC on all four HDR-IQA datasets, with a weighted average SRCC of 0.869 and PLCC of 0.873 versus 0.801/0.808 for HDR-VDP-3-Q.
- Luminance shift compensation improves weighted-average SRCC for Q-PSNR from 0.537 to 0.720 and for Q-DISTS from 0.708 to 0.869.
- In HDR novel view synthesis, HDR-NeRF† optimized with Q-MAE reaches HDR-VDP-3-Q 9.863 versus 7.023 for HDR-NeRF, and a 2AFC user study with 15 participants supports its perceptual gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed inverse-display-model HDR metric with DISTS instead of existing HDR quality metrics | HDR image quality assessment · Four HDR-IQA datasets | Weighted-average SRCC: 0.801 -> 0.869 | Helped |
| Add luminance shift compensation (ablation) | Q⋆DISTS · Four HDR-IQA datasets, weighted average | SRCC: 0.708 -> 0.869 | Helped |
| Add luminance shift compensation (ablation) | Q⋆MAE · Four HDR-IQA datasets, weighted average | SRCC: 0.525 -> 0.670 | Helped |
| Simplify HDR-NeRF by removing the tone mapper and optimize the HDR radiance field with the proposed metric | HDR-NeRF · HDR novel view synthesis, eight synthetic scenes | HDR-VDP-3-Q: 7.023 -> 9.863 | Helped |
| Replace the HDR-NeRF+ tone mapper with the proposed inverse display model optimization | HDR novel view synthesis · Eight synthetic scenes | PSNR: 27.358 -> 28.754 | Helped |
| Use Q⋆MAE instead of MAE as the HDR-NeRF† training loss (ablation) | HDR-NeRF† · HDR novel view synthesis, eight synthetic scenes | HDR-VDP-3-Q: 6.224 -> 9.863 | Helped |
| Use Q⋆MAE instead of PU21-MAE as the HDR-NeRF† training loss (ablation) | HDR-NeRF† · HDR novel view synthesis, eight synthetic scenes | HDR-VDP-3-Q: 9.717 -> 9.863 | Helped |
| Use Q⋆MAE instead of log-MAE as the HDR-NeRF† training loss (ablation) | HDR-NeRF† · HDR novel view synthesis, eight synthetic scenes | HDR-VDP-3-Q: 9.720 -> 9.863 | Helped |
| Use Q⋆MAE instead of µ-MAE as the HDR-NeRF† training loss (ablation) | HDR-NeRF† · HDR novel view synthesis, eight synthetic scenes | HDR-VDP-3-Q: 9.784 -> 9.863 | Helped |
| Use PU21-MAE rather than µ-MAE as the HDR-NeRF† training loss (ablation) | HDR-NeRF† · HDR novel view synthesis, eight synthetic scenes | HDR-VDP-3-Q: 9.784 -> 9.717 | No clear effect |
Surprising findings
- Encoding-based PU21 metrics degrade noticeably when applied to LDR images, whereas the proposed metrics preserve LDR-IQA capability.
Limitations
- Does not model reduced visual sensitivity at low absolute luminances, so predictions are identical for dark and bright displays.
- Relies on a fixed display model (1 to 200 cd/m2) and gain-offset-gamma parameters from prior work.
Experimental setup
Datasets: Narwaria2013 Valenzise2014 Zerman2017 UPIQ HDR-NeRF synthetic scenes
Benchmarks: Narwaria2013 Valenzise2014 Zerman2017 UPIQ
Metrics: SRCC PLCC PSNR SSIM HDR-VDP-3-Q
Baselines: NLPD HDR-VQM HDR-VDP-3-Q HDR-VDP-3-D PU-MAE PU21-PSNR PU21-SSIM PU21-LPIPS PU21-DISTS HDR-NeRF HDR-NeRF+
Builds on: HDR-VDP-3 HDR-NeRF PU21 encoding DISTS LPIPS
Closest papers by meaning
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.88)
- HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting (NeurIPS 2024 · similarity 0.87)
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.87)
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.86)
- LEDiff: Latent Exposure Diffusion for HDR Generation (CVPR 2025 · similarity 0.86)
- EvHDR-NeRF: Building High Dynamic Range Radiance Fields with Single Exposure Images and Events (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.