LIIFusion | ECCV 2026

Abstract

Multi-exposure fusion (MEF) expands the luminance range beyond what a single exposure can capture. Combining images taken at different exposure levels requires handling geometric differences while naturally merging their complementary brightness information. It often demands generative completion where details are missing. Diffusion-based generative methods address these challenges; however, they are computationally expensive and struggle to preserve fine structures in saturated regions. We propose LIIFusion, a coarse-to-fine framework that balances fusion quality and efficiency in generative MEF. The coarse stage performs low-resolution generative fusion, enhanced by adaptive exposure correction. The fine stage adapts a local implicit image function into a multi-exposure fusion function conditioned on high-resolution OE/UE sources and the coarse output.

What is Generative MEF?

When you take a photo, a camera can capture multiple frames at different exposures and integrate them to produce a visually pleasing image. This process is called multi-exposure fusion (MEF).

Conventional MEF aligns and blends information observed across differently exposed images. Under severe saturation, large motion, or occlusion, however, some scene content has no reliable correspondence.

Generative MEF uses a generative prior to synthesize such missing content instead of relying only on deterministic blending. Existing diffusion-based methods achieve strong fusion quality, but high-resolution inference requires expensive patch-wise sampling.

Conceptual illustration of multi-exposure fusion using over-exposed and under-exposed images to produce a fused result.

Main contributions

01 Coarse-to-fine Paradigm

Decouple luminance range integration and structural detail fusion.

02 INR as a Fusion Function

Reformulate LIIF from a super-resolution decoder into a multi-exposure conditional fusion function.

03 Efficiency and Performance

Achieve up to 3.7× faster inference than prior generative MEF while attaining state-of-the-art fusion quality.

LIIFusion teaser comparison highlighting quality and inference efficiency.

Method

Overview of the LIIFusion coarse-to-fine generative MEF pipeline. — LIIFusion replaces high-resolution patch-wise diffusion with low-resolution generative fusion and a LIIF-based high-resolution fine stage.

Adaptive Exposure Correction (AEC)

AEC attenuates severely over-exposed regions before fidelity guidance.

L^LR_diff = clip( L^LR_oe − L^LR_ue, 0, 1 )

W = (1 − α · L^LR_diff)^1/2.2

I′^LR_oe = I^LR_oe ⊙ W

The corrected OE image is used for fidelity guidance to enhance fusion quality at low resolution.

INR-based Fine Stage

Multi-resolution inputs—the low-resolution fused image and the high-resolution exposure images—are fused in a continuous feature domain using INR.

s = f_θ( [z_coarse, z_fine], [x, c])

Querying all target coordinates produces the final high-resolution fused image.

Quantitative results

Dynamic RealHDRV and UltraFusion Benchmark. Higher is better except Time.
Model	RealHDRV (50 scenes)					UltraFusion Benchmark (100 scenes)
Model	MUSIQ ↑	DeQA ↑	PAQ2PIQ ↑	HyperIQA ↑	Time ↓	MUSIQ ↑	DeQA ↑	PAQ2PIQ ↑	HyperIQA ↑	Time ↓
Defusion	56.38	3.2867	68.31	0.4838	2 min	60.11	3.3529	71.83	0.5440	6 min
MEF-LUT	62.42	3.2864	70.04	0.5020	4 sec	64.06	3.2859	71.80	0.5103	8 sec
HSDS-MEF	61.82	3.6045	71.14	0.5055	18 min	65.23	3.6662	73.77	0.5786	46 min
UltraFusion	67.54	3.8998	73.39	0.5834	101 min	68.40	4.0123	75.18	0.6214	203 min
LIIFusion (Ours)	69.52	3.8908	74.06	0.6175	27 min	70.19	3.9807	75.59	0.6467	59 min

Qualitative results

Qualitative comparison on ultra-high dynamic range scenes. — Ultra-high dynamic range scenes.

Qualitative comparison on dynamic motion scenes. — Dynamic motion scenes.

Additional results

Model	Best-choice rate ↑	Avg. rank ↓	Time ↓
MEF-LUT	3.79%	3.57	0.5 sec
HSDS-MEF	13.64%	2.89	2 min 27 sec
UltraFusion	21.21%	2.02	12 min 9 sec
LIIFusion (Ours)	61.36%	1.52	3 min 23 sec

Human preference study. Inference time is measured on an NVIDIA RTX A5000.

Comparison of local details in the under-exposed input, UltraFusion, and LIIFusion. — Effect of coarse-to-fine framework

Presentation

Citation

@article{han2026coarse,
  title={Coarse-to-fine Framework for Generative MEF via Implicit Neural Representation},
  author={Han, Sangmin and Kim, Jinho and Kim, Jinwoo and Kim, Dongyoung and Kim, Seon Joo},
  journal={arXiv preprint arXiv:2607.17611},
  year={2026}
}

LIIFusion: Coarse-to-fine Framework for Generative MEF via Implicit Neural Representation