Continuous Image Representation

Continuous image representation uses neural networks to represent images as functions of continuous coordinates, enabling resolution-independent processing and efficient manipulation. Current research focuses on developing computationally efficient architectures like latent modulated functions and attention-based networks, as well as improving training methods through techniques such as random weight factorization and mixture-of-expert training. This approach offers significant advantages in various applications, including image super-resolution, inpainting, harmonization, and 3D modeling, by reducing computational costs and improving generalization capabilities compared to traditional discrete representations.