Abstract
In this paper, we introduce novel lightweight generative adversarial networks, which can effectively capture long-range dependencies in the image generation process, and produce high-quality results with a much simpler architecture. To achieve this, we first introduce a long-range module, allowing the network to dynamically adjust the number of focused sampling pixels and to also augment sampling locations. Thus, it can break the limitation of the fixed geometric structure of the convolution operator, and capture long-range dependencies in both spatial and channel-wise directions. Also, the proposed long-range module can highlight negative relations between pixels, working as a regularization to stabilize training. Furthermore, we propose a new generation strategy through which we introduce metadata into the image generation process to provide basic information about target images, which can stabilize and speed up the training process. Our novel long-range module only introduces few additional parameters and is easily inserted into existing models to capture long-range dependencies. Extensive experiments demonstrate the competitive performance of our method with a lightweight architecture.
Introduction
Generative Adversarial Networks (GANs) have achieved remarkable success in image synthesis. However, standard architectures rely heavily on convolutional layers, which are constrained by local receptive fields. This limitation makes it difficult for models to capture long-range dependencies—such as the symmetry of a face or the structural consistency of a building—without stacking numerous layers or employing computationally expensive self-attention mechanisms. Such heavy architectures are often unsuitable for deployment on memory-constrained devices like mobile phones.
Methodology
The proposed architecture addresses these challenges through two main innovations: the Long-Range Module (LRM) and a metadata-driven generation strategy.
The Long-Range Module (LRM)
The LRM is designed to break the fixed geometric structure of traditional convolutions. It allows the network to dynamically adjust the number of focused sampling pixels and augment sampling locations through a free-form deformation of the sampling grid.
Unlike standard self-attention, which primarily focuses on positive correlations, the LRM can highlight negative relations between pixels. This capability acts as a regularization term, stabilizing the training process and preventing mode collapse. Furthermore, the LRM operates in both spatial and channel-wise directions with minimal parameter overhead.
Metadata-Driven Generation
To further stabilize and accelerate training, the authors introduce a metadata-driven strategy. Global features are extracted from target images using a pretrained VGG-16 network to provide the generator with high-level "metadata" (e.g., category and basic texture). This provides a structural prior that helps the generator converge faster without leaking fine-grained details that would lead to simple memorization.
Results and Discussion
The model was evaluated on benchmarks including FFHQ, CUB Birds, and ImageNet. The results demonstrate significant efficiency gains.
Efficiency and Speed
The proposed generator contains only 7.0M parameters, a substantial reduction compared to PGGAN's 23.3M. In terms of inference speed, the model can generate 100 images in just 0.83 seconds, making it approximately 20 times faster than PGGAN and twice as fast as SAGAN.
Image Quality
Despite its lightweight nature, the model achieves competitive Fréchet Inception Distance (FID) scores. On the CUB bird dataset, it significantly outperformed larger models, achieving an FID of 24.85 compared to PGGAN's 38.73. While slightly behind state-of-the-art heavyweights like StyleGAN on extremely complex datasets, it offers a superior trade-off for mobile and real-time applications.
Conclusion
Lightweight Long-Range GANs offer a promising path toward efficient, high-quality image synthesis. By replacing rigid convolutions with dynamic long-range modules and utilizing metadata for training stability, the researchers have created a model that captures the "big picture" without the computational burden of traditional high-capacity networks.
Summary updated