HDR Reconstruction Boosting with Training-Free and Exposure-Consistent Diffusion

Yo-Tin Lin; Su-Kai Chen; Hou-Ning Hu; Yen-Yu Lin; Yu-Lun Liu

cs.CVFeb 23, 2026

HDR Reconstruction Boosting with Training-Free and Exposure-Consistent Diffusion

Authors: Yo-Tin Lin, Su-Kai Chen, Hou-Ning Hu, Yen-Yu Lin, Yu-Lun Liu

Organizations: National Yang Ming Chiao Tung University · MediaTek Inc.

Abstract

Single LDR to HDR reconstruction remains challenging for over-exposed regions where traditional methods often fail due to complete information loss. We present a training-free approach that enhances existing indirect and direct HDR reconstruction methods through diffusion-based inpainting. Our method combines text-guided diffusion models with SDEdit refinement to generate plausible content in over-exposed areas while maintaining consistency across multi-exposure LDR images. Unlike previous approaches requiring extensive training, our method seamlessly integrates with existing HDR reconstruction techniques through an iterative compensation mechanism that ensures luminance coherence across multiple exposures. We demonstrate significant improvements in both perceptual quality and quantitative metrics on standard HDR datasets and in-the-wild captures. Results show that our method effectively recovers natural details in challenging scenarios while preserving the advantages of existing HDR reconstruction pipelines. Project page: https://github.com/EusdenLin/HDR-Reconstruction-Boosting

Figures & tables

Figure 2: Overview of our training-free HDR reconstruction pipeline. Given an input LDR image (EV0), we generate bracketed LDR images using an existing HDR reconstruction method. Our iterative pipeline then enhances these results through (1) an inpainting stage guided by exposure and condition maps, (2) HDR merging and alignment of the generated content, and (3) a compensation stage to ensure physical consistency. The top-right inset shows the progressive refinement of the over-exposed regions through our pipeline stages. The upper sky region in the aligned case shows the clear alignment across EV images. The lower sky region in the compensation and refined cases demonstrates our algorithm’s ability to hallucinate realistic over-expose regions with plausible intensity and texture. The upper part of the sky region in the aligned case shows the clear alignment across EV images. The lower parts of the sky region in the compensation and refined cases show the ability of our algorithm to hallucinate realistic over-expose regions with reasonable intensity and texture.

Figure 3: Limitations arising from naively combining indirect HDR reconstruction methods and over-exposed regions inpainting. Independent inpainting of each EV bracket, without cross-EV alignment, can introduce ghosting artifacts in the merged HDR result. These artifacts stem from inconsistencies in the generated content, becoming apparent after merging the independently inpainted exposures using Debevec’s method.

Figure 4: Motivation for luminance compensation in HDR reconstruction. The top row shows the existence of inpainted intensities below the lower bound of over-exposed regions may lead to unreasonable inverse CRF estimation and misaligned results. The bottom row demonstrates our intensity compensation approach, which ensures bounded intensities, resulting in proper inverse CRF estimation and exposure-aligned reconstruction that matches the input LDR.

Figure 5: Inpainting pipeline in our method. The depth-conditioned inpainting pipeline with scheduled strength is able to generate reasonable and consistent content across different EVs and iterations. The scheduled SDEdit strength enables us to balance detail preservation with creative generation. In the early iterations, we allow the model to explore a broader range of plausible scene details, while in later iterations, we retain the refined details from previous steps and only update pixels that do not meet the physical constraints.

Figure 6: Compensation pipeline in our method. Our pipeline ensures proper exposure relationships through iterative refinement. ( Top ) Overview of how inpainted LDRs are combined with luminance residuals to produce compensated results after tone-mapping. ( Bottom ) Visualization over four iterations shows inpainted results, decreasing residuals, and shrinking mask regions. The process maintains proper luminance lower bounds while focusing refinement on problematic areas through selective masking, preventing CRF estimation issues.

		Non reference											Ref.-based (vs. GT)
Dataset	Method	BRISQUE ↓		NIQE ↓		NIMA ↑		MUSIQ ↑		CLIP-IQA ↑		PU21 ↓	KID ↓		HDR- ↑
												PIQE			VDP-3
	TMO	RH’s	KK’s	RH’s	KK’s	RH’s	KK’s	RH’s	KK’s	RH’s	KK’s	–	RH’s	KK’s	–
VDS	SingleHDR [ 25 ]	75.73	73.08	14.49	14.25	4.021	4.352	2.032	0.295	2.063	0.318	90.16	183.9	205.1	7.5610
	SingleHDR + Ours	64.82	60.34	12.17	11.17	4.056	4.419	2.151	0.301	2.159	0.324	78.35	186.2	202.8	8.8604
	GlowGAN [ 50 ]	79.02	77.49	14.54	14.25	4.008	4.209	1.890	0.342	1.900	0.346	93.79	256.7	275.4	6.0815

Table 1: Quantitative comparisons. Our algorithm consistently improves state-of-the-art methods on both VDS [ 21 ] and HDR-Eye [ 33 ] datasets without additional training.

Figure 7: Qualitative comparisons of HDR reconstruction methods. Our approach enhances various baseline methods (indicated by “+ Ours ”) across diverse scenes, including outdoor architecture, indoor lighting, and urban landscapes. While existing methods often produce unrealistic content or artifacts in over-exposed regions, our method generates more natural details and lighting patterns while maintaining structural consistency without training. Results show improved reconstruction of sky details, cloud patterns, and lighting effects across different baseline methods. The top row shows the main difference between the baseline results and hallucinated results. Though we can generate a similar texture and color tone, the generated content and ground truth might not be considered the same by the KID score and HDR-VDP-3.

Figure 8: Inpainting diversity results. By leveraging the diffusion prior, our inpainting pipeline gains the ability to generate diverse scenes with different lighting conditions and sky appearances from the same input LDR image.

Method	NIQE ↓		NIMA ↑		PU21-PIQE ↓
TMO	RH’s	KK’s	RH’s	KK’s	–
CEVR [ 7 ]	14.33	13.48	4.044	4.421	90.44
CEVR + Ours (SDXL)	13.71	12.97	4.100	4.456	86.12
CEVR + Ours (SD turbo)	12.36	11.60	4.150	4.482	78.46

Table 2: Ablation study on diffusion backbones. Our algorithm improves state-of-the-art methods on the VDS dataset [ 21 ] with different diffusion backbones without additional training.

Figure 9: SDEdit strength scheduling ablation study. Low strength, e.g . , $0.6$ , limits content generation in over-exposed regions, while high strength, e.g . , $1.0$ , causes inconsistency between iterations. Our scheduling gradually decreases noise strength to balance content generation with detail refinement and consistency across all the four iterations.

Figure 10: Luminance compensation ablation study. ( Top ) Without compensation, the inpainted content may be unrealistically dark. ( Bottom ) Our compensation adjusts pixel values to maintain physically plausible brightness levels, as shown by decreasing residuals across iterations.

Components (RH’s TMO)	NIQE $\downarrow$	NIMA $\uparrow$	PU21-PIQE $\downarrow$
Baseline (CEVR)	14.33	4.044	90.44
w/o Merge & Comp.	13.57	4.045	85.15
w/o Merge	13.60	4.021	84.68
w/o Comp.	13.39	4.028	85.07
Ours	13.71	4.100	86.12

Table 3: Quantitative ablation studies. We conducted a quantitative evaluation on the VDS dataset and demonstrated that our approach yields improvements in non-reference image quality metrics.

Figure 11: Limitation. Our method is restricted by the estimated inverse CRF. When the estimated inverse CRF is unreasonable, our results suffer from color shifting as well.

HDR Imaging

High dynamic range imaging captures or reconstructs scene brightness beyond standard images, preserving detail in shadows and highlights.

See all 46 papers

ExpoCM: Exposure-Aware One-Step Generative Single-Image HDR Reconstruction
May 4, 2026Aoyu Liu, Zhen Liu, Ziyi Wang +3Image Enhancement HDR Imaging
University of Electronic Science and Technology of China
DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models
Apr 7, 2026Zhengming Yu, Li Ma, Mingming He +11HDR Imaging HDR Video Reconstruction
Texas A&M University · Eyeline Labs · Netflix
GMODiff: One-Step Gain Map Refinement with Diffusion Priors for HDR Reconstruction
Dec 18, 2025Tao Hu, Weiyu Zhou, Yanjie Tu +4Diffusion Models for Image Restoration Latent Diffusion Models
Northwestern Polytechnical University, Xi’an 710072, China · Xi’an University of Architecture and Technology, Xi’an 710055, China · The Shenzhen Research Institute of Northwestern Polytechnical University, Shenzhen 518057, China

May 4, 2026J/K move · Enter open · S save

ExpoCM: Exposure-Aware One-Step Generative Single-Image HDR Reconstruction

Aoyu Liu, Zhen Liu, Ziyi Wang, Dian Chen+2

University of Electronic Science and Technology of China

Image Enhancement HDR Imaging Image Reconstruction One-Step Generative Modeling Consistency Models Diffusion Models

Open paper AskPDF arXiv

Single-image HDR reconstruction aims to recover high dynamic range radiance from a single low dynamic range (LDR) input, but remains highly ill-posed due to detail saturation in over-exposed regions and noise amplification in under-exposed areas. While recent diffusion-based approaches offer powerful generative priors, they often overlook the exposure-dependent nature of the degradation and incur substantial computational costs from iterative sampling. To address these challenges, we propose ExpoCM, a novel one-step generative HDR reconstruction framework that reformulates HDR reconstruction as a Probability Flow ODE (PF-ODE) and constructs exposure-aware consistency trajectories via exposure-dependent perturbations. Specifically, a soft exposure mask is first constructed to separate the LDR image into over-, under-, and well-exposed regions. Based on this partition, region-conditioned consistency trajectories are designed to hallucinate saturated details, suppress noise in dark regions, and preserve reliable structures within a single, distillation-free inference step. To further enhance perceptual quality, we introduce an Exposure-guided Luminance-Chromaticity Loss in the CIE~

\text{L}^*\text{a}^*\text{b}^*

space, which assigns exposure-aware weights to luminance and chromaticity components, effectively mitigating brightness bias and color drift. Extensive experiments on the HDR-REAL, HDR-EYE, and AIM2025 benchmarks demonstrate that ExpoCM achieves state-of-the-art fidelity and perceptual accuracy, while enabling over 400

\times

and 20

\times

faster inference compared to DDPM (1000 steps) and DDIM (50 steps), respectively.

Abstract

HDR Imaging

Explore similar work

ExpoCM: Exposure-Aware One-Step Generative Single-Image HDR Reconstruction

DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models

GMODiff: One-Step Gain Map Refinement with Diffusion Priors for HDR Reconstruction