Multi-Garment Capability in Digital Fashion

• Multi-garment capability is the ability to represent, control, and process multiple garment layers simultaneously, addressing occlusion, layering, and distinct garment identities.
• Generative modeling architectures utilize parallel garment encoders, adaptive localization, and disentangled conditioning to enable realistic synthesis and editing of composite outfits.
• Robotic and physical systems apply multi-garment approaches for enhanced segmentation, collision handling, and manipulation, yielding significant quantitative improvements in occlusion-aware performance.

Multi-garment capability refers to the algorithmic, model-architectural, and systems-level ability to represent, control, and process multiple garments, garment types, or garment layers simultaneously within physical, virtual, or robotic workflows. It is a central methodological advance in recent computational garment research. Multi-garment capability encompasses compositional garment modeling, cross-layer interaction (e.g., occlusion, draping), independent control of garment identity, and the disentangled synthesis, editing, or manipulation of more than one article of clothing in a coherent manner. The domain spans generative modeling (image synthesis and editing), robotic manipulation, digital wardrobe reconstruction, and parametric 3D and pattern design.

1. Problem Definition and Conceptual Evolution

The core challenge motivating multi-garment capability is that realistic garment dressing, try-on, modeling, or manipulation in both virtual and physical domains inherently involves several articles of clothing, each with distinct class, geometry, material, and spatial relationships. Single-garment approaches cannot address combinatorial composition, occlusion reasoning, or per-garment control. Early efforts, such as Multi-Garment Net, introduced a layered garment-by-garment 3D parameterization to separate body and clothing geometry, enabling transfer, retargeting, or removal of individual layers (Bhatnagar et al., 2019). In generative modeling, subsequent advances focused on diffusion-based methods with parallel, decoupled garment encodings, plug-and-play attention fusion, and explicit occlusion-aware feature gating, establishing the paradigm for robust multi-garment synthesis and editing in virtual try-on, character customization, and pattern generation (Liu et al., 2024, Li et al., 2024, Yu et al., 20 Jan 2026).

Multi-garment capability now structurally underpins three technical subfields:

• Image and Video Synthesis: Simultaneous conditioning and synthesis from multiple reference garment images, free combination and layout, and occlusion-aware rendering for photorealistic try-on and editing.
• Physical and Robotic Manipulation: Category- and instance-aware perception and action planning for multiple garments in clutter, folding, and retrieval pipelines; explicit modulation of affordances by garment class and configuration (Chen et al., 11 Mar 2025, Li et al., 4 Mar 2026).
• Digital Pattern and Wardrobe Modeling: Composition, parsing, and editing of multi-garment outfits or modular designs via parametric or tokenized (panel-wise) representations, enabling cross-type outfit recovery, transfer, and simulation (Bian et al., 2024, Ackermann et al., 5 Jun 2025, Li et al., 2 Apr 2025, Lin et al., 13 Oct 2025).

2. Generative Modeling Architectures for Multi-Garment Synthesis

Contemporary generative models typically rely on latent diffusion frameworks in which multi-garment capability is realized through explicit modularization of the garment encoding and fusion pipelines:

• Parallel Garment Encoders and Feature Fusion: Models such as Multi-Garment Customized Model Generation (Liu et al., 2024) instantiate a garment encoder ("G_ϕ"), a parallel trainable UNet with shared weights, which processes each reference garment image independently to extract detailed multiscale feature maps $f_{i, l}$ per layer. These features are injected into the backbone UNet at each cross-attention block through decoupled "attention addition": each garment’s cross-attention is computed separately, and the resulting attention maps $A_i$ are summed, as

$$h_l' = h_l + A_\text{text} + \gamma \sum_{i=1}^N A_i,$$

where $\gamma$ modulates fusion strength. This approach mitigates information collision and preserves independent garment properties (texture, silhouette).

• Adaptive Localization and Occlusion Modeling: AnyDressing (Li et al., 2024) advances the pipeline with adaptive garment-specific LoRA adapters in attention modules, an instance-level garment localization loss that constrains each garment’s attention heatmap to its physical region (mask $M_k$), and garment-enhanced high-frequency texture regularization. GO-MLVTON (Yu et al., 20 Jan 2026) introduces a Garment Occlusion Learning module, producing a spatial attention map $A$ that gates (masks) inner-garment features to prevent "bleed-through" beneath overlying layers, reinforced by an explicit supervised occlusion loss:

$$\mathcal{L}_\mathrm{OCC} = \big\|\,\varepsilon(x_{p_i}) - z_i^v\big\|_2,$$

ensuring correct suppression of occluded details.

• Disentangled Conditioning and Layout Control: Models such as M&M VTO (Zhu et al., 2024) and BootComp (Choi et al., 2024) inject multiple garment embeddings in parallel into UNet/Transformer blocks, optionally with text-driven layout attributes, and apply cross-attention to spatially warp and place each garment on the target subject. Identity, pose, or style guidance can be plugged in via auxiliary modules (e.g., ControlNet, IP-Adapter).
• Quantitative Performance: Multi-garment LDM-based solutions report significant improvement on FID, CLIP-I, LPIPS, and custom per-garment alignment metrics over single-garment and naive fusion baselines; e.g., FID improves from 24.5 (single garment) or 18.7 (attention concatenation) to 12.3 (proposed attention addition), with concurrent gains in text/image consistency (Liu et al., 2024, Li et al., 2024).

3. Representation and Interaction in Pattern, 3D, and Modular Domains

Multi-garment capability in non-image domains centers on compositional, panel-wise, or token-wise representations and their manipulation:

• Parametric and Panel-Based Meshes: Multi-Garment Net (Bhatnagar et al., 2019) and ISP (Li et al., 2023) decompose a dressed human into a stack of separate garment meshes or SDF-based panel assemblies, each parametrized independently and layered over a body template (e.g., SMPL). ISP enables per-layer collision correction, intersection avoidance via explicit energy penalty, and supports sequential fitting or editing of outer and inner layers.
• Pattern Code, JSON Schemas, and Generative Sewing Representations: ChatGarment (Bian et al., 2024) extends GarmentCode to “GarmentCodeRC,” allowing structured JSON representations containing per-garment type and parameter blocks ("upperbody_garment", "lowerbody_garment", or "wholebody_garment"), supporting both parsing from images/sketches and interactive, part-level editing.
• Token and Panel Slot Fusion: Vision-language approaches (VLG (Ackermann et al., 5 Jun 2025)) and sequence diffusion (GarmageNet (Li et al., 2 Apr 2025)) parse and generate variable-length panel/garment slots through attention over garment-type tokens, with independent regression heads for each slot.
• Modular, ILP-Based Physical Garment Construction: Refashion (Lin et al., 13 Oct 2025) formulates garment assembly as an integer linear programming problem over reconfigurable modules, enabling physical interchange, resizing, and style mutation from a fixed library, thereby realizing combinatorial multi-garment creation and reuse.

4. Robotic and Manipulation Systems for Multi-Garment Environments

Robotic perception and manipulation pipelines now explicitly systematize multi-garment inputs, reasoning, and control.

• Segmentation, Reasoning, and Action Planning: GarmentPile++ (Li et al., 4 Mar 2026) couples state-of-the-art visual segmentation (SAM2) with vision-LLMs (VLMs) to decompose complex garment piles into discrete mask-labeled regions, then selects and manipulates single garments as per user language instructions, followed by affordance-driven grasp identification and post-grasp checks for single-garment retrieval.
• Hierarchical Task Decomposition: MetaFold (Chen et al., 11 Mar 2025) decomposes multi-garment folding into language-guided trajectory generation and a low-level action prediction module (ManiFoundation) trained across categorical variations. Explicit multi-category dataset design enables reliable generalization, e.g., the system achieves 0.88–0.97 success rates across shirts, pants, etc.
• Systematic Evaluation: Both GarmentPile++ and MetaFold present controlled ablation studies: multi-stage pipelines critically depend on segmentation accuracy, affordance estimation, dual-arm cooperation, and closed-loop re-planning. Removing these components sharply degrades multi-garment handling performance.

5. Multi-Garment Composition, Occlusion, and Conflict Handling

A central technical challenge is modeling interactions (occlusion, layering, collision, visual coherence) among garment layers:

• Attention Decoupling and Additive Fusion: Attention addition (as opposed to concatenation or naive blending) computes independent cross-attention maps for each garment and sums them, ensuring information is not overwritten and that the influence of each garment is additive (Liu et al., 2024). Gating vectors and per-category modulations further control interactions.
• Occlusion-Aware Feature Refinement: GO-MLVTON (Yu et al., 20 Jan 2026) computes spatial attention masks to soft-gate latent features from inner garments, suppressing regions occluded by outer garments. Supervised losses reinforce the removal of “hidden” information, and a Layered Appearance Coherence Difference (LACD) metric targets quantitative evaluation of cross-layer boundary consistency.
• Geometric Constraints and Intersection Losses: Physical and 3D modeling approaches constrain SDF or mesh-based fields with covering or collision penalties (e.g., $\mathcal{L}_\mathrm{cov}$ in LGN (Aggarwal et al., 2022) or $E_\text{collision}$ in ISP (Li et al., 2023)) to guarantee intersection-free garment stacks and stratified layer ordering.

6. Modular, Plug-and-Play, and Extensible Frameworks

A defining property of leading multi-garment solutions is modularity—architecturally and computationally:

• Plug-in Garment Encoders: Both Multi-Garment Customized Model Generation (Liu et al., 2024) and AnyDressing (Li et al., 2024) implement garment encoding modules as drop-in components compatible with external control modules (ControlNet, IP-Adapter, LoRA adapters), facilitating extensibility to new garment domains, conditioning signals, or additional tasks like pose, identity, or textual control.
• Fine-tuning and Personalization: Disentangled architecture in M&M VTO (Zhu et al., 2024) allows per-individual, low-memory fine-tuning (6MB per subject) for identity retention in multi-garment try-on.
• Modular Pattern and Simulation Tools: Refashion (Lin et al., 13 Oct 2025) includes an end-to-end digital design suite for physically grounded, mix-and-match garment assembly and simulation, backed by an ILP-based optimizer for module assignment.

7. Evaluations, Limitations, and Prospects

Benchmarking demonstrates that multi-garment frameworks outperform single-garment baselines on both aggregate and per-garment metrics (FID, LPIPS, CLIP-I, LACD, success rate in retrieval/manipulation). Notably:

• Quantitative gains are reported in image synthesis (e.g., FID 12.3 vs. 24.5 (Liu et al., 2024); LPIPS, CLIP-I in (Li et al., 2024)), 3D/intersection metrics (max penetration depth near zero in LGN (Aggarwal et al., 2022)), and robotic success rates (>0.9 in simulation and real-world for both folding and retrieval (Chen et al., 11 Mar 2025, Li et al., 4 Mar 2026)).
• Failure modes cluster around severe occlusion, densely overlapping garments, or insufficient training data for extreme compositions.
• Stated future directions include hierarchical grouping for scaling to $N>3$ garments, integration of garment segmentation masks, dynamic spatial gating, physics-aware regularization, and compositional language control at finer granularity.

A plausible implication is that as data resources, modular system designs, and cross-task transfer improve, multi-garment capability will define the foundation for the next generation of virtual try-on, robotic manipulation, and digital fashion design platforms. The documented approaches collectively set the agenda for extensible, open-ended multi-garment modeling, reasoning, and control (Liu et al., 2024, Li et al., 2024, Yu et al., 20 Jan 2026, Chen et al., 11 Mar 2025).