Core Concepts
TexTile introduces a differentiable metric to evaluate texture tileability, enhancing texture synthesis methods.
Abstract
The content introduces TexTile, a novel differentiable metric designed to quantify the tileable properties of textures. It addresses the limitations of existing perceptual metrics in evaluating texture synthesis algorithms. The article covers the development process, model design, training regime, dataset collection, and various applications of TexTile in texture synthesis algorithms. It also includes results from benchmarking different texture synthesis methods and demonstrates how TexTile can be used as a loss function to generate tileable textures while maintaining or improving overall image quality.
Directory:
Introduction
Existing methods focus on general texture quality but lack explicit analysis of intrinsic repeatability.
TexTile Development
Formulated as a binary classifier using architectural modifications.
Model Design and Training
Architecture combines convolutional and attention-based models.
Dataset Collection
Comprehensive dataset of tileable and non-tileable textures.
Evaluation
Ablation study on network architecture design and data augmentation policies.
Applications of TexTile
Alignment, repeating pattern detection, and transforming generative models into tileable texture synthesis algorithms.
Results
Quantitative comparison between different texture synthesis algorithms using reference and no-reference metrics.
Conclusion & Future Work
Stats
For each column, we show tiled versions of textures with (top) and without (bottom) tiling artifacts.