ExpandDiff: Dynamic Range Expanding Diffusion for Single-Image HDR Reconstruction

Mehmet Emre andıran; Zhuoqian Yang; Liying Lu; Mathieu Salzmann; Sabine Süsstrunk

cs.CVSep 30, 2026

ExpandDiff: Dynamic Range Expanding Diffusion for Single-Image HDR Reconstruction

Authors: Mehmet Emre andıran, Zhuoqian Yang, Liying Lu, Mathieu Salzmann, Sabine Süsstrunk

Organizations: School of Computer and Communication Sciences, EPFL, Switzerland

Abstract

Single-image HDR reconstruction requires inferring missing detail while preserving the visible content of an LDR image. Differences in sensor dynamic range and exposure cause LDR images to lose varying amounts of information in shadows and highlights. We present ExpandDiff, a conditional diffusion pipeline that jointly reconstructs clipped shadows and highlights. To account for this variation, we introduce Dynamic Clipping Synthesis (DCS), which randomly samples shadow and highlight clipping percentiles when constructing training inputs from HDR targets. A pixel-space diffusion model guided by spatially-adaptive normalization then predicts perceptually encoded HDR through a bounded output head, reconstructing both clipping directions in one sampling trajectory. On the SI-HDR benchmark, ExpandDiff variants improve HDR reconstruction accuracy by 3.43 dB in PU21-PSNR over the strongest evaluated competing method, and by 7.34 dB under two-sided clipping. The code and supplementary material are available at https://memreandiran.github.io/expanddiff/.

Figures & tables

Figure 1: ExpandDiff pipeline. Dynamic Clipping Synthesis varies the clipping levels of the LDR guidance during training, while PU21 encoding defines the HDR prediction domain. Spatial guidance conditions the diffusion network, and inverse encoding yields linear HDR.

$\mathrm{C_{95}}$
	PSNR $\uparrow$		VSI $\uparrow$		VDP-3 $\uparrow$		PIQE	FID
Method	gain	CRF	gain	CRF	gain	CRF	$\downarrow$	$\downarrow$
ExpandDiff-B	22.61	30.15	.9866	.9897	0 9.06	0 9.46	28.22	8.76
ExpandDiff-P	21.41	28.65	.9801	.9867	0 8.77	0 9.35	27.32	9.99
LDR input	20.47	27.93	.9756	.9826	0 8.44	0 9.15	28.15	10.00
ExpandNet	19.17	28.09	.9743	.9840	0 8.37	0 9.12	28.60	10.65

Table 1: Results on 181 SI-HDR references under two input conditions. PSNR, VSI, and PIQE use PU21; VDP-3 is in JOD. All scores include brightness alignment; CRF additionally applies the benchmark’s tone-and-color correction. FID denotes FID-R. Best values per condition are bold and second-best values underlined.

Panel of Figure 2 — Figure 2: Highlight reconstruction on two $\mathrm{C_{95}}$ inputs with $8.3\%$ and $6.9\%$ clipped highlights. Red marks highlight clipping. Each example occupies two rows. Reconstructions use gain-only alignment.

Panel of Figure 3 — Figure 3: Joint shadow and highlight reconstruction on two $\mathrm{C_{p}}$ inputs, clipped over $20.0\%$ and $20.5\%$ of the frame. Red and blue mark clipped highlights and shadows, respectively. Each example occupies two rows. Reconstructions use gain-only alignment.

	PSNR $\uparrow$		VSI $\uparrow$		VDP-3 $\uparrow$		PIQE	FID
Method	gain	CRF	gain	CRF	gain	CRF	$\downarrow$	$\downarrow$
ExpandDiff-P	31.21	33.52	.9913	.9917	0 9.62	0 9.68	39.31	0 4.78
PU21, no $\tanh$	30.20	33.08	.9905	.9910	0 9.59	0 9.67	39.46	0 4.77
linear, $\tanh$	29.21	31.38	.9890	.9897	0 9.51	0 9.60	40.04	0 6.03
linear, no $\tanh$	26.84	30.34	.9861	.9880	0 9.46	0 9.61	38.16	0 5.56

Table 2: PU21 encoding and the bounded head on

\mathrm{C_{p}}

. All variants use the same data and 150k training steps. Metric definitions and emphasis follow Table 1 .

Explore similar work

Single-Shot HDR Recovery via a Video Diffusion Prior
May 12, 2026Chinmay Talegaonkar, Jinshi He, Christopher McKenna +1High Dynamic Range Video Restoration
University of California San Diego · Creare LLC
GMODiff: One-Step Gain Map Refinement with Diffusion Priors for HDR Reconstruction
Dec 18, 2025Tao Hu, Weiyu Zhou, Yanjie Tu +4High Dynamic Range Diffusion Priors
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
DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models
Apr 7, 2026Zhengming Yu, Li Ma, Mingming He +11High Dynamic Range Video Diffusion Models
Texas A&M University · Eyeline Labs · Netflix

May 12, 2026, cs.CVJ/K move · Enter open · S save

Single-Shot HDR Recovery via a Video Diffusion Prior

Chinmay Talegaonkar, Jinshi He, Christopher McKenna, Nicholas Antipa

University of California San Diego · Creare LLC

High Dynamic Range Video Restoration Video Diffusion Models Exposure

Open paper AskPDF arXiv

Recent generative methods for single-shot high dynamic range (HDR) image reconstruction show promising results, but often struggle with preserving fidelity to the input image. They require separate models to handle highlights and shadows, or sacrifice interpretability by directly predicting the final HDR image. We address these limitations by re-casting single-shot HDR reconstruction as conditional video generation and fusing the generated frames into an HDR image. We finetune a video diffusion model to generate an exposure bracket, conditioned on a low dynamic range (LDR) input. We fuse this image bracket using per-pixel weights predicted by a light-weight UNet. This formulation is simple, interpretable, and effective. Rather than directly hallucinating an HDR image, it explicitly reconstructs the intermediate exposure stack and fuses it into the final output. Our method eliminates the need for separate models across exposure regimes and produces HDR reconstructions with high input fidelity. On quantitative benchmarks, we outperform state-of-the-art generative baselines with comparable model capacity on several reconstruction metrics. Human evaluators further prefer our results in 72% of pairwise comparisons against existing methods. Finally, we show that this input-conditioned sequence generation and fusion framework extends beyond HDR to other image reconstruction tasks, such as all-in-focus image recovery from a single defocus-blurred input.

Abstract

Explore similar work

Single-Shot HDR Recovery via a Video Diffusion Prior

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

DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models