Research map / Video / Video generation
Single Shot HDR Recovery via a Video Diffusion Prior
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.
Recasts single-shot HDR reconstruction as input-conditioned exposure-bracket generation by fine-tuning Stable Video Diffusion, then fuses the generated LDR frames with a light-weight per-pixel-weight Fusion UNet.
Problem
Single-shot HDR reconstruction from one LDR image is ill-posed; existing generative methods hallucinate content, need separate shadow/highlight models, or directly predict HDR with limited interpretability and input fidelity.
Approach
Treat an exposure bracket as a video with monotonically increasing exposure: fine-tune a video diffusion model conditioned on the input LDR to generate the bracket, then fuse frames with learned per-pixel weights supervised in PU-21 space.
Key results
- Ours reaches Q*-PSNR 22.36 vs 21.41 for BracketDiff and 21.24 for X2HDR-SD2.1, and Q*-LPIPS 0.289 vs 0.335 for X2HDR-SD2.1, at about 15s per image versus about 6 min for BracketDiff.
- In a 20-participant user study with 1920 pairwise comparisons, the method was preferred in 72.0% of trials overall.
- Fusion UNet with predicted brackets scores Q*-PSNR 22.36, above Mertens (18.87) and Debevec (15.23); an oracle with GT LDR brackets reaches 44.84, so the generated brackets are the main bottleneck.
Surprising findings
- Exposure-compensating predicted brackets had negligible effect on the final HDR score, indicating global EV drift is not the limiting factor.
- Fusion UNet underperformed Debevec merging on brackets produced after inference-time optimization to fixed 1-EV spacing, attributed to distribution shift.
Limitations
- Fine text and high-frequency details are degraded by the SVD VAE latent compression.
- Slightly behind baselines when input already contains well-captured shadows and highlights.
- Runtime (about 15s) is far slower than non-generative feedforward methods.
Experimental setup
Datasets: Fairchild HDRPS RawHDR SI-HDR NYUv2 LEDiff HDR sources
Benchmarks: SI-HDR raw2hdr test set (96 scenes) NYUv2 (50 test scenes, AIF)
Metrics: Q*-PSNR Q*-SSIM Q*-MAE Q*-LPIPS Q-JOD user preference win rate bracket consistency deviation
Baselines: LEDiff X2HDR-SD2.1 BracketDiff HDRCNN SingleHDR Mertens fusion Debevec merging
Builds on: Stable Video Diffusion EDM Mertens fusion Debevec merging PU-21 X2HDR BracketDiff
Closest papers by meaning
- Generating Content for HDR Deghosting from Frequency View (CVPR 2024 · similarity 0.91)
- Self-Supervised High Dynamic Range Imaging with Multi-Exposure Images in Dynamic Scenes (ICLR 2024 · similarity 0.89)
- EvHDR-GS: Event-guided HDR Video Reconstruction with 3D Gaussian Splatting (AAAI 2025 · similarity 0.88)
- High Dynamic Range Novel View Synthesis with Single Exposure (ICML 2025 · similarity 0.88)
- HDRFlow: Real-Time HDR Video Reconstruction with Large Motions (CVPR 2024 · similarity 0.88)
- Color Event Enhanced Single-Exposure HDR Imaging (AAAI 2024 · similarity 0.88)
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.