Physics Guided Diffusion Model Based Exposure Interpolation for High Dynamic Range Imaging

Zhengguo Li; Chaobing Zheng

cs.CVOct 27, 2026

Physics Guided Diffusion Model Based Exposure Interpolation for High Dynamic Range Imaging

Authors: Zhengguo Li, Chaobing Zheng

Organizations: Institute of Advanced Intelligence and Computing, A*STAR, Singapore, 138632 · the school of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, China

Abstract

Ghosting artifacts are an open problem for popular multi-shot based high dynamic range (HDR) imaging. An effective solution is to capture two large-exposure-ratio (LER) standard dynamic range (SDR) images simultaneously via a single-shot. Brightness order reversal (BOR) artifacts are generated when the two LER images are directly fused together. It is crucial to interpolate an image with an exposure time in between the exposure times of two LER images. In this paper, a novel physics-guided diffusion model based exposure interpolation algorithm is proposed to address the BOR artifacts. An initial image is interpolated by a physics-driven component, and the residual high-frequency details are generated by a new diffusion model based component. The new diffusion model is based on the correlation analysis of generative models and includes a new predicted target to reduce the variation of the Pearson correlation between the noisy data and the target, as well as the amplification factor of the fitting error. Compared with existing neural augmented exposure interpolation algorithms on top of convolution neural networks, the proposed exposure interpolation algorithm has an inherent advantage to generate more high frequency information. Experiment results validate the proposed exposure interpolation algorithm.

Figures & tables

Fig. 1: Overview of the proposed neural augmented exposure interpolation algorithm on top of a new diffusion model. Through reducing the variation of $\Psi_{\mathbf{e}_{t},\omega_{t}}$ as well as $|\Phi_{1}(t^{\prime},t)|$ and $|\Phi_{2}(t^{\prime},t)|$ , the fitting error of the neural network is reduced in the reverse process. Subsequently, the quality of the interpolated image is improved.

\mathcal{L}(\theta)=\left\|g_{\theta}^{(1)}(e_{t},f_{m}(Z_{1},Z_{7}),t)-\omega_{t}^{(1)}\right\|_{2}^{2}

Algorithm 1 Neural Augmented Exposure Interpolation Diffusion Model (Training)

e_{t^{\prime}}=\Phi_{1}^{(1)}(t^{\prime},t)e_{t}+\Phi_{2}^{(1)}(t^{\prime},t)\hat{\omega}_{t}^{(1)}

Algorithm 2 Neural Augmented Exposure Interpolation (Inference)

Fig. 2: $|\Phi_{1}(t^{\prime},t)|$ , $|\Phi_{2}(t^{\prime},t)|$ and $\Psi_{\mathbf{e}_{t},\omega_{t}}$ of two different choices of $\omega_{t}$ in equations ( 31 ) and ( 44 ) with $T$ as 5. $|\Phi_{2}^{(1)}(t^{\prime},t)|$ is smaller than $|\Phi_{2}^{(1)}(t^{\prime},t)|$ and the variation of $\Psi_{\mathbf{e}_{t},\omega_{t}^{(1)}}$ is smaller than the variation of $\Psi_{\mathbf{e}_{t},\omega_{t}^{(2)}}$ .

Fig. 3: Comparison of different physics-driven exposure interpolation algorithms. (a) are the inputs, (b), (c) and (d) show the results obtained by the physics-driven exposure interpolation algorithms in [ 18 , 19 ] , and our adopted physics-driven component, respectively. (e) represents the ground truth.

Methods	Data	PSNR $\uparrow$	SSIM $\uparrow$	FSIM $\uparrow$	LPIPS $\downarrow$	Entropy $\uparrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	HyperIQA $\uparrow$
Algorithm in [ 18 ]	[ 19 , 10 ]	26.65	0.9114	0.9761	0.0905	6.4661	0.4834	58.996	65.07
Algorithm in [ 19 ]	[ 19 , 10 ]	28.07	0.9142	0.98	0.0901	6.5903	0.4836	58.438	65.38
Our	[ 19 , 10 ]	29.79	0.919	0.9823	0.0848	6.691	0.4727	59.0875	65.46
Algorithm in [ 18 ]	[ 34 ]	24.4	0.8633	0.9519	0.1268	7.0101	0.4972	61.7405	65.36
Algorithm in [ 19 ]	[ 34 ]	24.93	0.8702	0.957	0.1192	7.0443	0.496	61.8327	65.44
Our	[ 34 ]	25.87	0.8848	0.9631	0.1083	7.0817	0.4998	61.8573	65.53

TABLE I: Objective comparison among the physics-driven exposure interpolation algorithms in [ 18 , 19 ] and our adopted physics-driven exposure interpolation component.

Target	PSNR $\uparrow$	SSIM $\uparrow$	FSIM $\uparrow$	LPIPS $\downarrow$	Entropy $\uparrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	HyperIQA $\uparrow$
Eq. ( 31 )	33.55	0.9489	0.9900	0.0441	6.8237	0.4702	63.40	67.23
Eq. ( 44 )	32.77	0.9429	0.9883	0.0492	6.8181	0.4704	61.89	66.94

TABLE II: Objective comparison between two different choices of

\omega_{t}

in equations ( 31 )(the proposed one) and ( 44 ) (one in [ 30 ] ) with

T

as 5 using date sets in [ 19 , 10 ] .

Fig. 4: Comparison of training curves for two different choices of $\omega_{t}$ in equations ( 31 ) and ( 44 ) with $T$ as 5.

$\alpha$	PSNR $\uparrow$	SSIM $\uparrow$	FSIM $\uparrow$	LPIPS $\downarrow$	Entropy $\uparrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	HyperIQA $\uparrow$
0.02	33.55	0.9489	0.9900	0.0441	6.8237	0.4702	63.40	67.23
0.1	33.44	0.9493	0.9899	0.0442	6.8139	0.4681	63.28	67.13
1.0	33.12	0.9447	0.9893	0.0454	6.8212	0.4775	62.77	66.91

TABLE III: Objective comparison between different choices of

\alpha

in equations ( 32 ) with

T

as 5 on [ 19 , 10 ] .

Fig. 5: Comparison of neural augmented exposure interpolation algorithms in [ 19 , 20 ] and the proposed neural augmented exposure interpolation algorithm. (a) Input images with different exposures. (b) Result of [ 19 ] . (c) Result of [ 20 ] . (d) Result of the proposed algorithm. The red circles mark the regions shown enlarged on the right of each result. The proposed algorithm produces sharper results with clearer details, especially in the highlighted foliage region.

Methods	Data	PSNR $\uparrow$	SSIM $\uparrow$	FSIM $\uparrow$	LPIPS $\downarrow$	Entropy $\uparrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	HyperIQA $\uparrow$
Algorithm in [ 19 ]	[ 19 , 10 ]	32.07	0.9362	0.9859	0.0527	6.7896	0.4541	58.80	66.19
Algorithm in [ 20 ]	[ 19 , 10 ]	33.41	0.9443	0.9894	0.0472	6.8137	0.4648	60.42	66.75
Our	[ 19 , 10 ]	33.55	0.9489	0.9900	0.0441	6.8237	0.4702	63.40	67.23
Algorithm in [ 19 ]	[ 34 ]	25.38	0.8822	0.9582	0.1239	7.1253	0.4455	60.64	65.51
Algorithm in [ 20 ]	[ 34 ]	26.62	0.8830	0.9613	0.1228	7.1333	0.4744	61.39	66.36
Our	[ 34 ]	26.71	0.8922	0.9621	0.1207	7.1406	0.4745	64.26	67.17

TABLE IV: Comparison of neural augmented exposure interpolation algorithms in [ 19 , 20 ] and the proposed neural augmented exposure interpolation algorithm.

Metrics	Data	GGIF [ 41 ]	MSF [ 18 ]	DeepFuse [ 14 ]	MEFNet [ 42 ]	UltraFusion [ 7 ]	DiFuser [ 43 ]	Proposed
MEF-SSIM	[ 19 , 10 ]	0.8717	0.8782	0.8579	0.8662	0.8834	0.7806	0.9172
MEF-SSIM	[ 34 ]	0.9126	0.9277	0.9104	0.9010	0.9409	0.8081	0.9484
TMQI	[ 19 , 10 ]	0.8328	0.8448	0.8219	0.8185	0.8365	0.8507	0.8768
PSNR	[ 34 ]	14.48	18.25	16.23	13.30	20.24	15.56	21.94
SSIM	[ 34 ]	0.7755	0.8119	0.7679	0.7429	0.8335	0.6884	0.8404

TABLE V: Objective comparison among the resultant MEF algorithm with 6 state-of-the-art exposure fusion algorithms with inputs as two LER images

Z_{1}

and

Z_{7}

(

\uparrow

: larger is better).

Fig. 6: Subjective comparison of seven exposure fusion algorithms with inputs as two LER images $Z_{1}$ and $Z_{7}$ . These two sets of differently exposed images are from the data set in [ 19 , 44 ] . All images in the set are given in the top two rows. BOR and halo artifacts, color distortion, loss of details in highlight/shadow regions, and reduced global contrast are possible issues for exposure fusion algorithms.

HDR Imaging

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

See all 47 papers

HDR Reconstruction Boosting with Training-Free and Exposure-Consistent Diffusion
Feb 23, 2026Yo-Tin Lin, Su-Kai Chen, Hou-Ning Hu +2HDR Imaging
National Yang Ming Chiao Tung University · MediaTek Inc.
Neural Block-wise Auto-exposure for Single-Shot High Dynamic Range Imaging
Oct 9, 2026Zhengguo Li, Chaobing Zheng, Wei WangHDR Imaging Computational Imaging
Institute of Advanced Intelligence and Computing (IAIC), Agency for Science, Technology and Research (A*STAR), Singapore, 138632 · the school of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, China · Advanced Remanufacturing and Technology Centre (ARTC), Agency for Science, Technology and Research (A*STAR)
Single-Shot HDR Recovery via a Video Diffusion Prior
May 12, 2026Chinmay Talegaonkar, Jinshi He, Christopher McKenna +1Image Enhancement HDR Imaging
University of California San Diego · Creare LLC

Feb 23, 2026J/K move · Enter open · S save

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

Yo-Tin Lin, Su-Kai Chen, Hou-Ning Hu, Yen-Yu Lin+1

National Yang Ming Chiao Tung University · MediaTek Inc.

HDR Imaging

Open paper AskPDF arXiv

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

Abstract

HDR Imaging

Explore similar work

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

Neural Block-wise Auto-exposure for Single-Shot High Dynamic Range Imaging

Single-Shot HDR Recovery via a Video Diffusion Prior