Single-image 3D reconstruction has made rapid progress, but standalone garments remain a hard case. Their thin-shell geometry, open boundaries, and layered structures are poorly served by existing datasets and evaluation methods, which limits systematic comparison. The new MagicGarment benchmark directly targets this gap.

The benchmark provides 1,518 professionally constructed garment assets across 17 composite garment types, along with aligned multimodal annotations and a unified evaluation protocol. This gives researchers a common ground for testing and comparing methods specifically on garment reconstruction rather than on general objects.

To demonstrate the benchmark's utility, the authors also build a garment-aware reference baseline by adapting a pretrained image-to-3D backbone at two levels. Grounded Semantic Refinement (GSR) supplies garment-focused visual conditioning, while Flow-Time-Route LoRA (FTR-LoRA) adjusts the frozen generative backbone using flow-timestep, token-routing, and expert-load information. Experiments show both the benchmark's usefulness and the effectiveness of this adaptation, though the paper is a short 5-page conference-style contribution and leaves broader comparisons to future work. (One source; no conflicting claims.)