Generative Convolution Layer for Image Generation
This addresses the challenge of enhancing image generation quality in GANs with minimal hardware cost and without architectural changes, though it appears incremental as it builds on existing convolution methods.
The paper tackles the problem of improving GAN performance by introducing a generative convolution (GConv) method that selects and combines kernels based on latent vectors to produce latent-specific features, resulting in significant boosts in image quality metrics such as FID and IS scores, e.g., improving FID from 35.13 to 29.76 and IS from 20.23 to 22.64 on tiny-ImageNet.
This paper introduces a novel convolution method, called generative convolution (GConv), which is simple yet effective for improving the generative adversarial network (GAN) performance. Unlike the standard convolution, GConv first selects useful kernels compatible with the given latent vector, and then linearly combines the selected kernels to make latent-specific kernels. Using the latent-specific kernels, the proposed method produces the latent-specific features which encourage the generator to produce high-quality images. This approach is simple but surprisingly effective. First, the GAN performance is significantly improved with a little additional hardware cost. Second, GConv can be employed to the existing state-of-the-art generators without modifying the network architecture. To reveal the superiority of GConv, this paper provides extensive experiments using various standard datasets including CIFAR-10, CIFAR-100, LSUN-Church, CelebA, and tiny-ImageNet. Quantitative evaluations prove that GConv significantly boosts the performances of the unconditional and conditional GANs in terms of Inception score (IS) and Frechet inception distance (FID). For example, the proposed method improves both FID and IS scores on the tiny-ImageNet dataset from 35.13 to 29.76 and 20.23 to 22.64, respectively.