AfriVTON-Bench: Evaluating the Robustness of Virtual Try-On Models to African Textile Patterns and Garment Structures

Virtual Try-On (VTON) systems now produce photorealistic results, but their evaluation rests on VITON-HD and DressCode, both curated from Western e-commerce platforms with lightly textured, body-congruent garments and predominantly lighter skin tones. African textiles stress what those corpora omit: Ankara, Kente and Aso-Oke carry high-frequency repeating motifs that strain warping and synthesis, while Agbada and Boubou have voluminous, non-fitted silhouettes. We present AfriVTON-Bench, a benchmark of 111 African garment images spanning 16 categories across 7 countries, paired with a 50-image Western control set from VITON-HD and 15 person images of African subjects labelled with Monk Skin Tone values. We run zero-shot inference with two architecturally distinct models, FASHN-VTON and LEFFA, yielding 1,012 person-garment outputs scored by garment-region and person-region perceptual metrics. Both models reconstruct person identity far more faithfully than garment texture (g/p 3.63x and 4.27x) and converge on near-identical garment distortion despite different designs. African garments lose 9.4 to 9.6 gSSIM points against Western controls, with Kente (0.460) and Ankara (0.474) falling 17 to 19 points below the Western control (0.640), while the structurally simpler Habesha Kemis (0.658) and Moroccan Kaftan (0.640) match it. Degradation therefore tracks garment visual complexity, not African origin. Three failure modes recur across both architectures: pattern dissolution, drape collapse, and tonal drift.