Neural fields are an emerging paradigm that represent data as continuous functions parameterized by neural networks. Despite many advantages, neural fields often have a high training cost, which prevents a broader adoption. In this paper, we focus on a popular family of neural fields, called sinusoidal neural fields (SNFs), and study how it should be initialized to maximize the training speed. We find that the standard initialization scheme for SNFs---designed based on the signal propagation principle---is suboptimal. In particular, we show that by simply multiplying each weight (except for the last layer) by a constant, we can accelerate SNF training by 10$\times$. This method, coined _weight scaling_, consistently provides a significant speedup over various data domains, allowing the SNFs to train faster than more recently proposed architectures. To understand why the weight scaling works well, we conduct extensive theoretical and empirical analyses which reveal that the weight scaling not only resolves the spectral bias quite effectively but also enjoys a well-conditioned optimization trajectory.

Image-to-image translation with generative adversarial networks (GANs) has been extensively studied in recent years. Among the models, StarGAN has achieved image-to-image translation for multiple domains with a single generator, whereas conventional models require multiple generators. However, StarGAN has several limitations, including the lack of capacity to learn mappings among large-scale domains; furthermore, StarGAN can barely express small feature changes. To address the limitations, we propose an improved StarGAN, namely SuperstarGAN. We adopted the idea, first proposed in controllable GAN (ControlGAN), of training an independent classifier with the data augmentation techniques to handle the overfitting problem in the classification of StarGAN structures. Since the generator with a well-trained classifier can express small features belonging to the target domain, SuperstarGAN achieves image-to-image translation in large-scale domains. Evaluated with a face image dataset, SuperstarGAN demonstrated improved performance in terms of Fréchet Inception distance (FID) and learned perceptual image patch similarity (LPIPS). Specifically, compared to StarGAN, SuperstarGAN exhibited decreased FID and LPIPS by 18.1% and 42.5%, respectively. Furthermore, we conducted an additional experiment with interpolated and extrapolated label values, indicating the ability of SuperstarGAN to control the degree of expression of the target domain features in generated images. Additionally, SuperstarGAN was successfully adapted to an animal face dataset and a painting dataset, where it can translate styles of animal faces (i.e., a cat to a tiger) and styles of painters (i.e., Hassam to Picasso), respectively, which explains the generality of SuperstarGAN regardless of datasets.