Reproduce “Adversarial Generation of Continuous Images” on OpenResearch

About this paper

Ivan Skorokhodov, Savva Ignatyev, Mohamed Elhoseiny

This research from King Abdullah University of Science and Technology introduces an implicit neural representation (INR) based generative adversarial network (INR-GAN) for high-resolution continuous image synthesis, addressing challenges of hypernetwork instability and computational expense. The method achieves competitive FID scores of 16.32 on FFHQ 1024x1024, demonstrating inherent properties like out-of-the-box superresolution and a stronger geometric prior.

Key insights

• The use of sigmoid activation in Factorized Multiplicative Modulation (FMM) is crucial for stabilizing hypernetwork training and achieving high-quality continuous image generation.
• INR-based decoders inherently offer desirable functionalities like out-of-the-box superresolution, meaningful image-space interpolation, and a strong geometric prior without explicit training or additional supervision.
• Multi-scale INRs significantly mitigate computational costs, making high-resolution continuous image generation feasible and even more efficient in MACs than some pixel-based counterparts.

Summary by alphaXiv.