OmniFashion

OmniFashion is a research project from Wuhan University and Harbin Institute of Technology focused on generalist fashion intelligence via multi-task vision-language learning. The team introduced FashionX, a 1.03M-outfit dataset with hierarchical GPT-4.1-annotated style attributes across thirteen subtasks, and trained the 3B-parameter OmniFashion model that currently sets state-of-the-art on fashion understanding benchmarks. Academic fashion-AI projects of this scope matter to the industry because their datasets and model weights frequently become the foundation layers that commercial vendors fine-tune rather than train from scratch. The use of GPT-4.1 as an annotation engine for FashionX is a methodological choice that ties the dataset's quality ceiling to a proprietary model, a dependency that will shape how the benchmark ages as both the annotation model and the evaluation tasks evolve.

First mentioned: April 2026