Generative AI for Fashion and Body Shape Embeddings
Open Access DepositedGenFab
Online fashion return rates range from 30% to 40%, with poor garment fit driving the majority of returns. Existing recommendation systems rely on garment size as a proxy for fit and rarely account for variation in body shape.This praxis introduces GenFab, a garment–body compatibility prediction framework derived from the ViBE architecture. The proposed approach replaces ViBE’s text-mined binary attribute vector with a 384-dimensional GPT-4o semantic embedding, while preserving the original four-branch multilayer perceptron structure. Two configurations were evaluated
GenFab-I, using a 14-dimensional SMPL body representation, and GenFab-II, using a compact 4-dimensional anthropometric representation (height, bust, waist, hips). Architecture, training protocol, and evaluation procedures were held constant to isolate the impact of semantic attribute enrichment. Results show that GenFab-II improved ranking AUC by 16.4% and prediction accuracy by 13.1% for tops in the person-unseen evaluation scenario (p ≈ 0, Bonferroni-corrected α = 0.00417). Additional improvements were observed in the both-unseen scenario, with gains of 6.6% in AUC and 1.5% in accuracy. All three hypotheses were supported. These findings indicate that vision–language semantic garment representations can substantially improve garment–body compatibility prediction, providing a practical mechanism for reducing fit-related dissatisfaction and potentially lowering return rates in fashion e-commerce.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
CityJarman_gwu_0075A_17915.pdf | 2026-06-24 | Open Access |
|