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

ECCV 2026

Sangmin Han1,2 Jinho Kim1 Jinwoo Kim1 Dongyoung Kim1 Seon Joo Kim1

1Yonsei University   2AI Lab, CTO Division, LG Electronics

Teaser comparison of Inputs, HSDS-MEF, UltraFusion, and LIIFusion with zoomed-in regions and runtime/MUSIQ labels.
LIIFusion preserves natural generative fusion quality while recovering fine structures and reducing runtime compared with patch-wise generative MEF.

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 that recovers structure lost in saturated over-exposed areas. The fine stage adapts local implicit image functions for fusion, enabling resolution-agnostic, pixel-wise refinement capable of restoring fine detail.

New paradigm of Generative MEF

global generative synthesis followed by high-resolution implicit refinement.

Reformulation of INR

First usage of implicit function for image fusion.

Break limitation of using diffusion

3.7x speed-up over exisiting Generative MEF method while achieving SOTA performance.

Background

Conventional MEF Patch-wise Generative MEF Coarse-to-fine Generative MEF
Models CNN, Transformer Diffusion Diffusion, LIIF
Dynamic Range Limited (3-4 stops) Extended (9 stops) Extended (9 stops)
Motion Handling Static / Mild Dynamic Dynamic
Fusion Behavior Regressive Probabilistic Both
Speed Fast (~minutes) Slow (~hours) Fast (~minutes)

Method

LIIFusion pipeline comparing patch-wise generative MEF with the proposed coarse-to-fine generative MEF.
Patch-wise generative MEF requires many diffusion sampling passes. LIIFusion minimizes diffusion usage through low-resolution coarse fusion with adaptive exposure correction, then performs resolution-agnostic pixel-wise detail fusion with a LIIF-based fine stage.

Quantitative Results

Dynamic RealHDRV and UltraFusion Benchmark results. Higher is better except Time.
Model RealHDRV (50 scenes) UltraFusion Benchmark (100 scenes)
MUSIQ ↑DeQA ↑PAQ2PIQ ↑HyperIQA ↑Time ↓ MUSIQ ↑DeQA ↑PAQ2PIQ ↑HyperIQA ↑Time ↓
Defusion 56.383.286768.310.48382 min 60.113.352971.830.54406 min
MEF-LUT 62.423.286470.040.50204 sec 64.063.285971.800.51038 sec
HSDS-MEF 61.823.604571.140.505518 min 65.233.666273.770.578646 min
UltraFusion 67.543.899873.390.5834101 min 68.404.012375.180.6214203 min
Ours 69.523.890874.060.617527 min 70.193.980775.590.646759 min

Qualitative Results

Qualitative comparison on ultra high dynamic range scenes using MEF-LUT, HSDS-MEF, UltraFusion, and LIIFusion.
Ultra HDR scenes: LIIFusion maintains stable exposure fusion, structural detail, and natural color appearance.
Qualitative comparison on dynamic motion scenes with ghosting and occlusion handling.
Dynamic motion scenes: LIIFusion suppresses ghost artifacts while restoring missing structural details.