Direction-Aware GMADL Objective

• The paper introduces the GMADL framework which integrates explicit gradient orientation to achieve sharper edge consistency and richer texture details in image fusion and multi-objective tasks.
• It details a multi-scale, axis-wise gradient loss and a direction-oriented optimization model that employs soft regularization to interpolate between conflict-averse and preference-aligned descent directions.
• Empirical results demonstrate that models using this objective yield superior semantic detail and efficient convergence without altering underlying architectures.

A Direction-Aware GMADL (Gradient-Magnitude-and-Direction Loss) Objective formalizes the use of explicit directional supervision in gradient-based objectives, particularly in multi-task, multi-objective, and image modality fusion settings. Unlike traditional gradient-magnitude approaches, the direction-aware paradigm maintains and leverages gradient orientation, yielding sharper, semantically consistent feature boundaries and meaningful optimization trajectories. Two principal developments in this area are the direction-aware multi-scale gradient loss for image fusion (Yang et al., 15 Oct 2025) and the direction-oriented multi-objective optimization framework (Xiao et al., 2023), both of which integrate axis- or vector-level directionality into learning objectives.

1. Motivation for Direction-Aware Objectives

Standard gradient-magnitude losses and classical multi-objective optimization schemes collapse vector information and suppress axis- or modality-specific directional cues. In image fusion, this leads to ambiguous edge supervision and weak texture transfer. In multi-objective optimization, naïve averaging or Pareto tracing can obscure task-preferred descent directions, especially in the presence of conflicting gradients. Direction-aware objectives are motivated by the need to retain meaningful gradient directionality at the pixel or parameter level, thereby enhancing edge consistency, texture preservation, and interpretable multi-criteria descent (Yang et al., 15 Oct 2025, Xiao et al., 2023).

2. Mathematical Formulation

Image Fusion: Multi-Scale, Axis-Wise Gradient Loss

The Direction-Aware Multi-Scale Gradient Loss, denoted $\mathcal{L}_{\mathrm{GMADL}}$, supervises the horizontal and vertical Sobel gradient components independently at multiple spatial scales (Yang et al., 15 Oct 2025). For input images (infrared and visible, single-channel), gradients are extracted:

• $\nabla_x I$ and $\nabla_y I$ computed using Sobel kernels $K_x$ and $K_y$.
• At each spatial scale $s \in S$, images are bilinearly resampled, and gradients $\nabla_x(s), \nabla_y(s)$ are computed.
• At each pixel and axis, modality selection masks $M_x(s;h,w), M_y(s;h,w)$ pick the channel (visible or infrared) with the higher absolute gradient, preserving the sign.

The per-scale loss is

$$L_s = \sum_{h,w} |\nabla_x^f(s;h,w) - \nabla_x^{\mathrm{sel}}(s;h,w)| + |\nabla_y^f(s;h,w) - \nabla_y^{\mathrm{sel}}(s;h,w)|$$

Aggregating over scales (with per-scale weights $w_s$) yields

$$\mathcal{L}_{\mathrm{GMADL}} = \sum_{s\in\mathcal{S}} w_s L_s$$

where axiswise selection

$$\nabla_x^{\mathrm{sel}}(s;h,w) = M_x(s;h,w)\nabla_x^{\mathrm{vis}}(s;h,w) + (1 - M_x(s;h,w))\nabla_x^{\mathrm{ir}}(s;h,w)$$

and similarly for $\nabla_y^{\mathrm{sel}}$.

Multi-Objective Learning: Direction-Oriented Descent

Given $K$ smooth objectives $L_i:\R^m\to\R$ with gradients $g_i(x)$ in parameter space $\R^m$, and a reference combination $g_0(x)$, the direction-aware objective is formulated as

$$\max_{d\in\R^m} \min_{i=1,\dots,K} \langle g_i(x), d\rangle - \frac{1}{2}\|d\|^2 + \lambda \langle g_0(x), d\rangle$$

or equivalently as a dual minimization over the simplex: $\min_{w\in\W} \frac{1}{2} \|G(x)w + \lambda g_0(x)\|^2$ where $G(x) = (g_1(x),\dots,g_K(x)) \in \R^{m\times K}$. The regularization parameter $\lambda$ controls the neighborhood around $g_0$ in which the common descent direction $d$ is chosen (Xiao et al., 2023).

3. Directionality and Sign Preservation Mechanisms

In $\mathcal{L}_{\mathrm{GMADL}}$, directionality is maintained by axiswise supervision—each Sobel gradient component is independently aligned and its sign is preserved across scales and pixels. The modality with higher local absolute gradient is chosen per axis, propagating high-frequency, direction-specific features to the fused image at both fine and coarse resolutions. This is in contrast to prior magnitude-based losses, which lose directional information through absolute or squared aggregation (Yang et al., 15 Oct 2025).

For multi-objective learning, the direction-aware penalty $\lambda\langle g_0, d\rangle$ produces a soft neighborhood constraint around a user-specified preference direction, interpolating between pure multi-gradient descent (MGDA) and gradient descent on a reference linear combination. The descent is 'aware' of directional task preferences and conflict, avoiding degenerate compromise solutions (Xiao et al., 2023).

4. Aggregation and Integration into Training Objectives

Both image fusion and multi-objective frameworks incorporate the direction-aware objective as one component of a composite loss:

• In image fusion, $\mathcal{L}_{\mathrm{GMADL}}$ is combined with intensity reconstruction ($\mathcal{L}_{\mathrm{int}}$; typically $L_1$ on the fused image versus the per-pixel maximum of the visible and IR channels) and a structural similarity term ($\mathcal{L}_{\mathrm{SSIM}}$). The final training objective is

$$\mathcal{L}_{\mathrm{total}} = \lambda_{\mathrm{SSIM}}\mathcal{L}_{\mathrm{SSIM}} + \lambda_{\mathrm{int}}\mathcal{L}_{\mathrm{int}} + \lambda_{\mathrm{grad}}\mathcal{L}_{\mathrm{GMADL}}$$

with, for example, $$\lambda_{\mathrm{SSIM}}:\lambda_{\mathrm{int}}:\lambda_{\mathrm{grad}}=1.5:7:1.5$$ in ReCoNet-based experiments (Yang et al., 15 Oct 2025).

• In direction-aware multi-objective learning, the soft penalty produces an SGD-friendly, smooth landscape, and the formulation recovers classical MGDA at $\lambda=0$ and linear-combination descent as $\lambda\to\infty$ (Xiao et al., 2023).

5. Advantages over Traditional Magnitude-Based and Conflict-Averse Approaches

Direction-aware objectives offer several distinct benefits:

• Edge and texture fidelity: By directly supervising vector gradients and their signs, fused images exhibit sharper, better-aligned edges and richer texture transmission than those produced using magnitude-only losses (Yang et al., 15 Oct 2025).
• Semantic interpretability: Tasks and modalities with distinct directional information exert explicit influence, enabling finer-grained multi-objective optimization and image integration strategies.
• Smooth trade-off control: The regularization parameter $\lambda$ enables continuous interpolation between conflict-averse (worst-case decrease) and preference-aligned (average decrease) strategies, offering practical and theoretical flexibility (Xiao et al., 2023).
• Efficient implementation: Both the image fusion and MOO formulations admit computationally simple, scalable implementations. In MOO, the dual simplex reduction and soft constraint facilitate the use of unbiased stochastic gradient estimators.

6. Empirical Results and Practical Impact

Empirical evaluation with public benchmarks and open-source models has demonstrated the effectiveness of the direction-aware GMADL objective in image fusion. Qualitatively, models trained with this loss produce fused images with superior structure, edge, and semantic detail compared to those trained under previous losses. These improvements are achieved without modifying underlying model architectures or training protocols (Yang et al., 15 Oct 2025).

In multi-objective settings, the direction-oriented formulation provably converges to Pareto stationary points for fixed $\lambda$ and yields convergence to single-objective optima as $\lambda$ increases. Comparative studies report improved performance and robustness in multi-task supervised learning and reinforcement learning (Xiao et al., 2023).

7. Relationship to Related Methodologies

Direction-aware GMADL builds upon and generalizes prior conflict-averse and multi-gradient approaches. It subsumes the MGDA multi-gradient direction for $\lambda=0$ and recovers gradient descent on reference objectives in the infinite-penalty limit. Compared with hard-constraint methods such as CAGrad, which enforce strict proximity to a mean direction, the soft regularization strategy in direction-aware objectives enables superior flexibility and tractable stochastic optimization. This suggests broad applicability across structured prediction, multi-modal integration, and multi-objective policy learning (Xiao et al., 2023).