Neural Knitworks: Patched Neural Implicit Representation Networks

TL;DR

Neural Knitwork introduces a patch-based neural implicit model that improves image synthesis tasks like inpainting and super-resolution, achieving higher fidelity with fewer parameters than traditional CNNs.

Contribution

It is the first coordinate-based MLP designed specifically for image synthesis tasks, leveraging patch-based training and adversarial optimization.

Findings

Achieves higher fidelity in image synthesis tasks.

Uses 80% fewer parameters than CNN-based models.

Performs comparably in training time and quality.

Abstract

Coordinate-based Multilayer Perceptron (MLP) networks, despite being capable of learning neural implicit representations, are not performant for internal image synthesis applications. Convolutional Neural Networks (CNNs) are typically used instead for a variety of internal generative tasks, at the cost of a larger model. We propose Neural Knitwork, an architecture for neural implicit representation learning of natural images that achieves image synthesis by optimizing the distribution of image patches in an adversarial manner and by enforcing consistency between the patch predictions. To the best of our knowledge, this is the first implementation of a coordinate-based MLP tailored for synthesis tasks such as image inpainting, super-resolution, and denoising. We demonstrate the utility of the proposed technique by training on these three tasks. The results show that modeling natural images using patches, rather than pixels, produces results of higher fidelity. The resulting model requires 80% fewer parameters than alternative CNN-based solutions while achieving comparable performance and training time.