Pattern Analogies: Learning to Perform Programmatic Image Edits by Analogy

A novel method for programmatic edits on pattern images uses a pattern analogy and a learning-based generative model, avoiding inference of the underlying program. This approach, Analogical Editing, employs a domain-specific language, SPLITWEAVE, and a Latent Diffusion Model, TRIFUSER, to create a large synthetic training dataset for pattern edits, successfully generalizing beyond its training distribution on real-world patterns. The method surpasses previous methods in fidelity and generalization, offering a versatile tool for pattern manipulation. TRIFUSER, an architecture combining text-image and self-supervised features, excels in generating high-quality, analogically edited patterns, demonstrating superior performance in complex pattern editing and structure preservation.