FusionNet: Multi-model Linear Fusion Framework for Low-light Image Enhancement

TL;DR

FusionNet formulates a parallel, multi-model linear fusion for low-light image enhancement, uniting CIDNet (HVI), Retinexformer (sRGB), and ESDNet (CNN) under a single, train-together, test-time linear fusion. Grounded in Hilbert space theory, the method selects fusion weights $k_i$ to maximize projection onto the target space, mitigating network collapse and reducing training overhead, while preserving independence of the base models. The approach achieves state-of-the-art performance on NTIRE2025 and across LOL datasets, delivering robust improvements in PSNR and SSIM and competitive perceptual quality. The work demonstrates that linear, parallel fusion can efficiently combine complementary representations across color spaces to deliver high-quality LLIE with strong generalization for diverse, real-world scenes, and suggests extensions to dynamic weighting for further gains.

Abstract

The advent of Deep Neural Networks (DNNs) has driven remarkable progress in low-light image enhancement (LLIE), with diverse architectures (e.g., CNNs and Transformers) and color spaces (e.g., sRGB, HSV, HVI) yielding impressive results. Recent efforts have sought to leverage the complementary strengths of these paradigms, offering promising solutions to enhance performance across varying degradation scenarios. However, existing fusion strategies are hindered by challenges such as parameter explosion, optimization instability, and feature misalignment, limiting further improvements. To overcome these issues, we introduce FusionNet, a novel multi-model linear fusion framework that operates in parallel to effectively capture global and local features across diverse color spaces. By incorporating a linear fusion strategy underpinned by Hilbert space theoretical guarantees, FusionNet mitigates network collapse and reduces excessive training costs. Our method achieved 1st place in the CVPR2025 NTIRE Low Light Enhancement Challenge. Extensive experiments conducted on synthetic and real-world benchmark datasets demonstrate that the proposed method significantly outperforms state-of-the-art methods in terms of both quantitative and qualitative results, delivering robust enhancement under diverse low-light conditions.

Figures (5)

• Figure 1: Comparison with recent SOTA methods in LOLv2-real lol_v2 dataset. We choose PSNR$\uparrow$ and SSIM$\uparrow$ for the measure metrics. The size of the circles indicates the FLOPs of each model. The result shows that proposed FusionNet achived the best performance among these methods.
• Figure 2: Four multi-model fusion strategies: (a) Serially Connect Network: Multiple networks are trained in a cascaded manner, but the probability of achieving optimal fitting in Hilbert space is low, making it difficult to converge to the best solution. (b) Multi-stage Serial Network: The parameters of the previous stage are frozen while training the next stage. However, this often leads to an increased number of iterations in later stages, making it more challenging to approach the optimal solution. (c) Parallel-stage Serial Network: Each model is trained separately in the earlier stages, and their outputs are concatenated before being fused by a new network. However, in Hilbert space, this method is fundamentally similar to (b) and does not resolve the issue of difficult convergence. (d) Linear Fusion Network: A simple yet efficient linear fusion operation is employed to enhance multi-model fusion performance, bringing the output space closer to the target space in Hilbert space.
• Figure 3: Visual comparison of the enhanced images yielded by five different SOTA methods snr_netcai2023retinexformerhou2024globalyan2025hvibai2024retinexmamba and proposed FusionNet on LOLv1 (top row), LOLv2-real (middle row), and LOLv2-synthetic (bottom row).
• Figure 4: Visualization of different fusion strategies of RetinexFormer cai2023retinexformer, CIDNet yan2025hvi, and ESDNet yu2022towards.
• Figure 5: Three-dimention visualization of different $k_i$s on PSNR and SSIM in LOLv2-real dataset.