Метод виявлення дублікатів workflow для LoRA-моделей на основі канонізації графових структур та структурного хешування

Abstract

This paper proposes a method for detecting duplicate and near-duplicate workflows used in LoRA-based image generation systems. Workflows exported in JSON format are modeled as directed graphs and transformed into canonical representations independent of node identifiers and interface metadata. The proposed approach combines structural normalization, topological sorting, cryptographic hashing (SHA-256) and SimHash-based similarity analysis. The method enables accurate detection of both exact duplicates and slightly modified workflow versions. Experimental evaluation demonstrates high precision and low false-positive rates in a simulated marketplace environment. The developed solution can be integrated into AI workflow marketplaces to ensure content uniqueness and protect intellectual property