Hybrid Weighting Strategy for Multi-Loss-Function of Aesthetics Aware GAN

Abstrak

Generative Adversarial Networks (GANs) have been widely extended with multiple loss functions to encourage the generation of images with desirable properties beyond realism, such as compositional balance and color aesthetics. The inclusion of multiple aesthetic losses introduces a multi-objective optimization setting in GAN training, where inappropriate loss weighting can undermine adversarial learning and compromise training stability. To address this issue, we propose a hybrid loss weighting strategy that anchors the adversarial loss while dynamically adjusting auxiliary aesthetic objectives. Specifically, Rule of Thirds loss and Colorfulness loss are integrated into a standard GAN framework, and three weighting schemes are systematically evaluated: static weighting, Gradient Normalization (GradNorm), and Dynamic Weight Average (DWA), including anchored and non-anchored variants. Experiments on CIFAR-10 demonstrate that aesthetic-oriented objectives consistently improve aesthetic scores without degrading generative quality. Moreover, the proposed anchored dynamic strategy, particularly GradNorm with adversarial anchoring, achieves the best overall trade-off between aesthetic enhancement, Fréchet Inception Distance (FID), and Inception Score (IS). These results confirm that controlled anchoring is essential for stable and effective multi-objective GAN optimization.

Identifier

• OpenAlex: https://openalex.org/W7151949796
• DOI: https://doi.org/10.1109/access.2026.3682107