Research map / Image generation / Image restoration
Generating Content for HDR Deghosting from Frequency View
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 LF-Diff, a diffusion model operating in a compact latent space that predicts low-frequency priors from LDR images, which guide a regression-based HDR deghosting network, achieving SOTA quality at ~10x lower inference time than DiffHDR.
Problem
Ghost-free HDR reconstruction from multiple LDR images with saturation and motion is hard; existing diffusion-based HDR methods generate full HDR images from noise, requiring many iterations and large models, which is inefficient.
Approach
Run the diffusion model in a highly compressed latent space to predict a low-frequency prior representation (LPR) from LDR inputs, rather than the full image, and inject this prior into a regression network (DHRNet) via cross-attention, with frequency-decomposed feature refinement blocks.
Key results
- On Kalantari's dataset LF-Diff reaches 44.76 PSNR-µ and 42.59 PSNR-L, exceeding DiffHDR by 0.65 dB and 0.86 dB.
- Inference takes 0.72 s per 1000x1500 image versus 7.53 s for DiffHDR, with 7.48M parameters versus 74.99M.
- Joint training of DM and DHRNet outperforms split training by 1.51 dB PSNR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LF-Diff instead of HyHDR | HDR reconstruction · Kalantari’s dataset | PSNR-µ: 44.64 -> 44.76 | Helped |
| use LF-Diff instead of HyHDR | HDR reconstruction · Hu’s dataset | PSNR-µ: 48.46 -> 48.74 | Helped |
| add ground-truth-derived LPR priors (ablation) | LF-Diff · Kalantari’s dataset | PSNR-L: 42.19 -> 50.89 | Helped |
| replace LPENet LPR prediction with DM prediction (ablation) | LF-Diff · Kalantari’s dataset | PSNR-µ: 44.45 -> 44.76 | Helped |
| jointly train DM and DHRNet instead of split training (ablation) | LF-Diff · Kalantari’s dataset | PSNR-L: 39.30 -> 42.59 | Helped |
| set reverse-process sampling step S to 5 instead of 10 (ablation) | LF-Diff · Kalantari’s dataset | PSNR-µ: 44.76 -> 39.76 | Hurt |
| increase reverse-process sampling step S from 10 to 20 (ablation) | LF-Diff · Kalantari’s dataset | PSNR-µ: 44.76 -> 44.73 | No clear effect |
Surprising findings
- Increasing the DDIM sampling steps beyond 10 gives no PSNR gain, so a small step count suffices in the latent space.
Limitations
- Complete elimination of ghosting is not achieved for severe motion, per the authors' qualitative results on unlabeled datasets.
- Evaluation of generalization relies only on qualitative comparisons on datasets without ground truth.
Experimental setup
Datasets: Kalantari dataset Hu dataset Sen et al. dataset (qualitative) Tursun et al. dataset (qualitative)
Benchmarks: Kalantari dataset Hu dataset Sen et al. dataset Tursun et al. dataset
Metrics: PSNR-µ SSIM-µ PSNR-L SSIM-L HDR-VDP-2 inference time
Baselines: Sen et al. patch-based Hu et al. patch-based Kalantari et al. DHDR AHDR NHDRRNet HDR-GAN APNT ADNet CA-ViT HyHDR DiffHDR
Builds on: DiffHDR DDIM AHDR Kalantari optical-flow HDR variance-preserving diffusion model
Closest papers by meaning
- Single Shot HDR Recovery via a Video Diffusion Prior (NeurIPS 2026 · similarity 0.91)
- Zero-Shot Structure-Preserving Diffusion Model for High Dynamic Range Tone Mapping (CVPR 2024 · similarity 0.90)
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.88)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.87)
- LEDiff: Latent Exposure Diffusion for HDR Generation (CVPR 2025 · similarity 0.87)
- Towards Real-World HDR Video Reconstruction: A Large-Scale Benchmark Dataset and A Two-Stage Alignment Network (CVPR 2024 · 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.