dev-tyta/Afri-VTON

Evaluating the Robustness of Virtual Try-On Models to African Textile Patterns and Garment Structures # AfriVTON-Bench **Evaluating the Robustness of Virtual Try-On Models to African Textile Patterns and Garment Structures** *Testimony Adekoya · Shiloh Oni* --- ## Summary State-of-the-art Virtual Try-On (VTON) models are trained on datasets dominated by Western casual wear. This work investigates whether these models generalise to garments with **high-frequency textile patterns** and **voluminous non-Western silhouettes** — properties that stress-test known weaknesses in current VTON pipelines. We proceed in two stages: 1. **Representation audit** — examine how existing VTON datasets (VITON-HD, DressCode, DeepFashion) represent garments with dense textile patterns and non-Western garment structures. 2. **Benchmark evaluation** — introduce AfriVTON-Bench and evaluate 4 SOTA VTON models zero-shot, measuring how synthesis quality degrades as pattern complexity and structural difference increase. African traditional attire — Ankara, Kente, Agbada, Boubou, and related garments — provides the primary test distribution because these garments systematically exhibit the visual properties that challenge current architectures. --- ## Research Questions **Q1 — Representation** Do existing VTON datasets sufficiently represent garments with dense textile patterns and non-Western silhouettes? **Q2 — Robustness** How does the performance of SOTA VTON models change when applied to garments with higher pattern complexity and different garment geometries? **Q3 — Failure modes** What types of errors occur? (pattern blurring, motif misalignment, garment shape distortion, texture drift) --- ## Why African Garments? African traditional textiles expose three specific weaknesses in current VTON pipelines: - **High-frequency patterns** — Ankara and Kente contain dense repeating motifs with strong geometric structure. Spatial warping modules and diffusion decoders frequently blur or hallucinate these during synthesis. - **Voluminous silhouettes** — Agbada and Boubou have loos …