MagicGarment: Benchmarking Single-Image Standalone 3D Garment Reconstruction
Single-image reconstruction of standalone garments remains under-supported by existing benchmarks, which mostly target clothed humans or rigid objects. MagicGarment introduces a dataset of 1,518 synthetic assets across 17 composite types, paired with aligned multimodal annotations and a unified evaluation protocol. The authors also propose a garment-aware baseline that adapts a frozen Hunyuan3D backbone via input-level semantic refinement (GSR) and a dynamic timestep- and routing-conditioned LoRA (FTR-LoRA), demonstrating that domain-specific adaptation improves over generic 3D priors on this benchmark.
The paper makes a timely, well-scoped contribution by formalizing standalone garment reconstruction and backing it with a dense, richly annotated benchmark. The proposed adaptation improves appearance and surface-normal metrics, but the overall geometric coverage (F-score) drops sharply relative to the backbone: the text notes the baseline achieves the best result on "9 of 13 metrics," yet its F-score of 9.0500 trails far behind Hunyuan3D-2.1 (14.0527) and TRELLIS.2 (15.2347), suggesting the adaptation prioritizes surface fidelity over structural completeness.
The benchmark design is convincing and fills a clear gap. The zero-shot evaluation in Table 2 reveals that no existing method dominates across geometry, appearance, and semantics, validating the need for a unified garment benchmark. The qualitative comparison supports that the baseline "more consistently preserves the garment silhouette, open boundaries, and local structural details across diverse examples." Additionally, Section 2.1 notes the dataset provides "aligned multi-view RGB images, masks, depth maps, normal maps, contours, sketches, textual descriptions, complete and partial point clouds, and 2D pattern-related information," which supports broad downstream use.
The dataset is entirely synthetic, created in CLO with artist-crafted patterns and 500 template bodies; the paper acknowledges this limitation by noting future work will extend to "more challenging real-world scenarios," but the current benchmark offers no real-image validation.
More critically, the garment-aware adaptation underperforms in geometric completeness: the baseline F-score is the lowest among adapted methods in Table 3, indicating FTR-LoRA may trade coverage for surface detail. The GSR mask-selection objective in Eq. 1 also depends on five hand-weighted terms ($w_q, w_u, w_b, w_\rho, w_s$) whose values and sensitivities are not reported or ablated.
The authors evaluate 11 methods under a unified protocol, which is fair and transparent. They appropriately caveat that AIpparel and ChatGarment "operate within predefined garment spaces without general texture generation." The quantitative claims are backed by Tables 2 and 3, though the semantic metrics rely on aggregated point-cloud similarities (ULIP/Uni3D), which can be brittle for thin, open shells. The comparison would be stronger with cross-dataset or real-world transfer experiments, as the current evidence is limited to in-domain synthetic evaluation.
Training details are only partially disclosed: the paper states the baseline is "trained using AdamW with a learning rate of $1\times 10^{-5}$ for 700K steps on four NVIDIA A30 GPUs. At inference, we use 50-step Euler ODE sampling for shape-latent generation." However, crucial hyperparameters for GSR (the weights in Eq. 1 and 2), LoRA ranks, and MoE controller dimensions are omitted. Because assets are professionally constructed in CLO with manual quality inspection, independent dataset reproduction is labor-intensive and poorly specified. The paper links a project page but does not confirm code or data release in the provided text, which are necessary to reproduce the baseline results.
Single-image 3D reconstruction has advanced rapidly, yet standalone garments remain challenging due to thin-shell geometry, open boundaries, and layered structures. Existing datasets and evaluation protocols provide limited support for systematic evaluation in this setting. We introduce MagicGarment, a benchmark for single-image standalone 3D garment reconstruction with 1,518 professionally constructed assets across 17 composite garment types, aligned multimodal annotations, and a unified evaluation protocol. We further establish a garment-aware reference baseline by adapting a pretrained image-to-3D backbone at two levels: Grounded Semantic Refinement (GSR) provides garment-focused visual conditioning, while Flow-Time-Route LoRA (FTR-LoRA) adapts the frozen 3D generative backbone using flow-timestep, token-routing, and expert-load information. Experiments demonstrate MagicGarment's benchmark utility and the effectiveness of garment-aware adaptation. Project page: https://1019538157.github.io/MagicGarment/.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.