EAFNet: Exposure-Adaptive Fusion for HDR Videos

• The paper proposes a dual-camera HDR video reconstruction system that uses a stable medium-exposure reference alongside low-/high-exposure frames to reduce temporal instability.
• It introduces an exposure-adaptive network featuring exposure-aware pre-alignment, reference-dominated asymmetric cross-feature fusion, and DWT-based multiscale reconstruction to enhance HDR quality.
• Empirical results show improved PSNR and temporal stability compared to traditional methods, validating the effectiveness of the dual-camera design and fusion strategy.

Searching arXiv for the requested EAFNet paper and closely related exposure-fusion work. Exposure-Adaptive Fusion Network (EAFNet) is an HDR video reconstruction model introduced in the context of a dual-camera system (DCS) designed to avoid the exposure fluctuation and flickering associated with alternating-exposure capture. In the formulation used by the method, one camera captures a consistent medium-exposure reference sequence, while the other captures non-reference low-/high-exposure sequences for information supplementation. EAFNet then reconstructs an HDR frame aligned to the stable reference by combining exposure-aware pre-alignment, reference-dominated cross-feature fusion, and DWT-based multiscale reconstruction (Zhang et al., 9 Jul 2025).

1. Problem setting and acquisition paradigm

EAFNet is defined for HDR video reconstruction in dynamic scenes from multiple LDR images captured at different exposures. The paper formulates HDR imaging as the problem of generating an HDR image $H$ from a set of LDR images $\{L_1,L_2,\ldots,L_n\}$ while aligning the output to a prescribed reference image. Its immediate motivation is the observation that alternating-exposure HDR video pipelines are temporally unstable because the reference frame changes exposure over time, producing inconsistent brightness, unstable reconstruction in highlight and shadow regions, and visible flickering (Zhang et al., 9 Jul 2025).

The proposed dual-camera system addresses this acquisition-level instability by separating reference capture from dynamic-range supplementation. The primary camera continuously captures a medium-exposure sequence $L_{reference}(t)$, while the secondary camera captures an alternating low-/high-exposure non-reference sequence $L_{non-reference}(t)$. The reference stream therefore remains exposure-consistent across time, and the supplementary stream provides shadow and highlight information that would be unavailable in a single fixed exposure. The paper explicitly links this design to reduced luminance fluctuation, reporting that Fig. 1 compares alternating-exposure methods with the stable-reference setup and shows luminance fluctuation exceeding 30 for alternating exposure but under 1 for the proposed approach (Zhang et al., 9 Jul 2025).

This acquisition strategy changes the reconstruction problem rather than eliminating it. The non-reference frames remain heterogeneous in exposure, are only roughly grouped by timestamps, and can exhibit parallax, temporal offset, object motion, occlusion, and local luminance mismatch. EAFNet is therefore built around three assumptions stated or implied by the paper: the medium-exposure frame is the geometrically and temporally most trustworthy input, the low-/high-exposure frames are necessary but risky supplementary observations, and temporal stability is obtained primarily from the stable reference stream rather than from recurrent temporal propagation. This suggests that EAFNet is best understood as a reference-anchored HDR fusion architecture rather than as a generic video sequence model (Zhang et al., 9 Jul 2025).

2. Network organization and input representation

The core input of EAFNet is a triplet of exposure-ordered RGB images,

$\{L_l,L_m,L_h\},$

where $L_l$, $L_m$, and $L_h$ denote low-, medium-, and high-exposure images, and $L_m$ is the reference image. Under the DCS setup, $L_m$ is the stable frame from the primary camera, while $\{L_1,L_2,\ldots,L_n\}$0 and $\{L_1,L_2,\ldots,L_n\}$1 are supplementary frames from the secondary camera. The target output is a reconstructed HDR image $\{L_1,L_2,\ldots,L_n\}$2, later assembled into a temporally stable HDR video by processing each reference frame in sequence (Zhang et al., 9 Jul 2025).

EAFNet is organized into three subnetworks: a pre-alignment subnetwork, an asymmetric cross-feature fusion subnetwork, and a restoration or reconstruction subnetwork. The data flow begins by augmenting each input exposure with two additional radiometric variants. First, gamma correction produces

$\{L_1,L_2,\ldots,L_n\}$3

with $\{L_1,L_2,\ldots,L_n\}$4. Second, Global Luminance Alignment (GLA) produces aligned luminance-domain images $\{L_1,L_2,\ldots,L_n\}$5, where $\{L_1,L_2,\ldots,L_n\}$6. The paper then defines the network input as

$\{L_1,L_2,\ldots,L_n\}$7

and describes this as “3-pairs of 9-channel tensor” (Zhang et al., 9 Jul 2025).

Initial $\{L_1,L_2,\ldots,L_n\}$8 convolutions extract exposure-specific features

$\{L_1,L_2,\ldots,L_n\}$9

with non-reference features $L_{reference}(t)$0 and reference features $L_{reference}(t)$1. These are processed in a multiscale manner by exposure-guided selection blocks, reference-dominated asymmetric cross-attention, and finally a DWT-based reconstruction hierarchy. The architecture therefore places the exposure triplet into three progressively more structured representations: image-domain triplets $L_{reference}(t)$2, branch-specific feature maps $L_{reference}(t)$3, and multiscale fused features that are decoded to the final HDR output (Zhang et al., 9 Jul 2025).

3. Pre-alignment subnetwork: luminance normalization and exposure-guided selection

The pre-alignment subnetwork is intended to make subsequent fusion exposure-aware before explicit cross-feature interaction begins. Its two components are Global Luminance Alignment and the Exposure-Guided Feature Selection Module (EFSM). The paper motivates this stage by noting that exposure changes distort color and luminance distributions, suppress shadow detail in low exposures, destroy highlight detail in high exposures, and reduce the reliability of attention maps if those maps are computed directly on heterogeneous inputs (Zhang et al., 9 Jul 2025).

GLA operates in the sRGB domain and maps non-reference images toward the luminance range of the reference. For $L_{reference}(t)$4, the paper defines

$L_{reference}(t)$5

while

$L_{reference}(t)$6

The paper states that this maps non-reference images to the luminance range of the reference image. It also notes a possible tension between the verbal description and the printed ratio, so the exact expression should be preserved as stated rather than silently altered. This suggests that GLA is a pragmatic luminance-conditioning step whose primary role is to reduce brightness inconsistency before exposure-guided modulation (Zhang et al., 9 Jul 2025).

EFSM incorporates explicit exposure metadata into feature selection. The relative low/high exposure values are defined with respect to the middle exposure: $L_{reference}(t)$7 with $L_{reference}(t)$8. At each scale, the non-reference feature $L_{reference}(t)$9 is passed through a shared-weight Exposure-Guided Feature Selection Block, while the reference feature $L_{non-reference}(t)$0 is passed through a shared-weight Feature Selection Block. Global average pooling, fully connected layers, and sigmoid activation produce feature-guided coefficients

$L_{non-reference}(t)$1

and the exposure input produces

$L_{non-reference}(t)$2

These are fused for the non-reference branch as

$L_{non-reference}(t)$3

yielding the modulated features

$L_{non-reference}(t)$4

This asymmetry is structurally important. The reference branch is modulated only by feature content, whereas the non-reference branch is modulated by both feature content and exposure metadata. The paper’s ablation results indicate that EFSM performs best only when GLA is also used; EFSM without GLA can degrade performance because brightness inconsistency makes exposure-guided modulation unreliable. The reported qualitative effect is that GLA + EFSM enhances shadow details in low-exposure features and suppresses overexposed regions in high-exposure features (Zhang et al., 9 Jul 2025).

4. Asymmetric cross-feature fusion and reference-dominated attention

The asymmetric cross-feature fusion subnetwork is the central fusion mechanism of EAFNet. Its defining claim is that the reference frame should dominate the fusion process because it is the most trustworthy source of geometry and temporal consistency, whereas the low-/high-exposure frames are informative but prone to parallax error, temporal mismatch, motion ghosting, and luminance inconsistency. The model therefore does not treat reference and non-reference features symmetrically (Zhang et al., 9 Jul 2025).

At each scale, Asymmetric Cross-Attention (ACA) computes patch-token attention using the reference feature as the dominant guide and the non-reference feature as the auxiliary source. With $L_{non-reference}(t)$5 denoting the unfolding of features into patch tokens, the paper defines

$L_{non-reference}(t)$6

and

$L_{non-reference}(t)$7

where $L_{non-reference}(t)$8, $L_{non-reference}(t)$9, and $\{L_l,L_m,L_h\},$0 are learnable projection matrices. The paper interprets this as a reference-dominated attention map because the reference feature is injected into the query term while the non-reference keys and values remain the complementary information source. The aligned non-reference feature $\{L_l,L_m,L_h\},$1 is then obtained by multiplying $\{L_l,L_m,L_h\},$2 and $\{L_l,L_m,L_h\},$3 and folding the result back into the spatial layout (Zhang et al., 9 Jul 2025).

Fusion is further stabilized by coarse-to-fine guidance. For $\{L_l,L_m,L_h\},$4, the paper defines

$\{L_l,L_m,L_h\},$5

where $\{L_l,L_m,L_h\},$6 is pixel-shuffle upsampling, $\{L_l,L_m,L_h\},$7 is a $\{L_l,L_m,L_h\},$8 convolution, and $\{L_l,L_m,L_h\},$9 is an operation not clearly specified in the text. The intent is nevertheless explicit: coarser-scale attention is upsampled and used to guide finer-scale alignment, so that robust coarse correspondences constrain finer local matching (Zhang et al., 9 Jul 2025).

The fused non-reference representation is then constructed as

$L_l$0

followed by a final spatial attention stage: $L_l$1 This final expression makes the asymmetry fully explicit. The non-reference feature is spatially weighted by an attention mask computed jointly with the reference feature, while the medium-exposure reference feature $L_l$2 is preserved directly. In effect, the supplementary features are admitted into the fused representation only after reference-guided spatial filtering. The fusion subnetwork is therefore not a generic cross-attention block; it is a reference-anchored gating mechanism tailored to dual-camera HDR reconstruction (Zhang et al., 9 Jul 2025).

5. DWT-based reconstruction and optimization objective

After asymmetric fusion, EAFNet reconstructs the HDR image through a DWT-based multiscale restoration architecture. The paper motivates this design by arguing that residual ghosting often appears as high-frequency double edges, local texture inconsistencies, and ringing near moving boundaries. Discrete Wavelet Transform separates features into frequency subbands, making it possible to correct high-frequency errors more explicitly than with a purely spatial decoder (Zhang et al., 9 Jul 2025).

At lower scales, the fused features are decomposed by DWT and each frequency component is corrected using multiple Learnable Bandpass Filter (LBF) modules before inverse wavelet reconstruction. At higher scales, DWT is still used, but the correction is simplified to a $L_l$3 convolution instead of multiple LBF modules, reflecting the paper’s claim that higher-scale features primarily encode more global structure and need less aggressive high-frequency correction. The multiscale outputs are finally merged as

$L_l$4

where $L_l$5, $L_l$6, and $L_l$7 denote features at different scales (Zhang et al., 9 Jul 2025).

Training is performed in the $L_l$8-law tone-mapped domain. The tone-mapping transform follows

$L_l$9

The final objective combines an $L_m$0 reconstruction term with a dilated advanced Sobel loss. The advanced Sobel loss is

$L_m$1

and the multi-dilation form is

$L_m$2

With $L_m$3 and $L_m$4, the total loss is

$L_m$5

The reported training configuration uses PyTorch, an NVIDIA RTX4090, $L_m$6 convolutions with 64 kernels, stride 1, zero padding, Adam optimization, an initial learning rate of $L_m$7, training termination at $L_m$8, crop size $L_m$9, batch size 16, and random rotation augmentation. The paper does not specify the exact number of epochs, the precise number of scales $L_h$0, the wavelet basis used for DWT/IWT, the internal structure of the LBF modules, or the exact patch size used by $L_h$1. This suggests that the code release is important for exact reproduction even though the high-level reconstruction logic is clear (Zhang et al., 9 Jul 2025).

6. Empirical performance, ablations, and research context

On public HDR reconstruction datasets, EAFNet is reported to achieve state-of-the-art or near-state-of-the-art results. When trained and tested on Kalantari’s dataset, it achieves PSNR-$L_h$2 $L_h$3, PSNR-L $L_h$4, SSIM-$L_h$5 $L_h$6, SSIM-L $L_h$7, and HDR-VDP-2 $L_h$8. On Prabhakar’s dataset, it achieves PSNR-$L_h$9 $L_m$0, PSNR-L $L_m$1, SSIM-$L_m$2 $L_m$3, SSIM-L $L_m$4, and HDR-VDP-2 $L_m$5. Cross-dataset evaluation is also reported to be strong, indicating robustness under exposure and motion variation (Zhang et al., 9 Jul 2025).

The most distinctive system-level result concerns temporal stability on self-captured dual-camera video. On the DCS setup, EAFNet used with MEF reconstruction reports luminance standard deviation $L_m$6 and t-SSIM $L_m$7, whereas alternating-exposure methods exhibit much larger luminance standard deviation, approximately $L_m$8–$L_m$9, and t-SSIM up to $L_m$0. This supports the paper’s central claim that temporal stability arises first from the stable-reference acquisition paradigm and then from reference-dominated fusion (Zhang et al., 9 Jul 2025).

The ablation studies make the internal division of labor explicit. Pre-alignment improves from a baseline PSNR-$L_m$1 of $L_m$2 to a full-model PSNR-$L_m$3 of $L_m$4, with GLA alone helping but EFSM without GLA potentially hurting. The ratio between feature-guided and exposure-guided modulation performs best at $L_m$5. In fusion, replacing ACA with standard cross-attention reduces PSNR-$L_m$6 from $L_m$7 to $L_m$8, and removing the guidance feature $L_m$9 reduces it to $\{L_1,L_2,\ldots,L_n\}$00. In reconstruction, removing multiscale structure or DWT also degrades performance, although the gains from the wavelet-domain component are more modest than those from pre-alignment and asymmetric fusion (Zhang et al., 9 Jul 2025).

Within the broader literature, EAFNet occupies a specific position rather than a generic one. FCNet addresses arbitrary-length exposure estimation, including both Single-Exposure Correction and Multi-Exposure Fusion, through Laplacian Pyramid decomposition and alternating fusion and correction (Liang et al., 2022). CRMEF addresses misalignment and efficiency through scene relighting, deformable alignment, detail repletion, and hardware-sensitive architecture search (Liu et al., 2023). The “Holistic Dynamic Frequency Transformer” unifies low-light enhancement, exposure correction, and multi-exposure fusion through Laplacian pyramid decomposition and frequency-domain attention (Shang et al., 2023). AFUNet formulates HDR reconstruction as cross-iterative alignment and fusion via deep unfolding, with Alignment-Fusion Modules alternating a Spatial Alignment Module and a Channel Fusion Module (Li et al., 30 Jun 2025). Relative to these methods, EAFNet is distinguished by its system-level dependence on a dual-camera stable-reference pipeline and by its combination of exposure-guided pre-alignment, reference-dominated asymmetric cross-feature fusion, and DWT-based reconstruction (Zhang et al., 9 Jul 2025).

The paper also implies several limitations. The method depends on dual-camera hardware, assumes a low/mid/high exposure triplet with the mid exposure as reference, and is designed for mild rather than extreme parallax and temporal offset. Several architecture details remain under-specified, including the number of scales, the patch size for tokenization, the exact operator $\{L_1,L_2,\ldots,L_n\}$01, and the wavelet basis. This suggests that EAFNet should be viewed not only as a neural architecture but as a capture-and-reconstruction system whose strongest contribution is the coupling of a stable-reference acquisition design with an explicitly exposure-adaptive fusion model (Zhang et al., 9 Jul 2025).