Research map / Image generation / Text-to-image diffusion
AnyDressing: Customizable Multi-Garment Virtual Dressing via Latent Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AnyDressing, a latent-diffusion framework (GarmentsNet + DressingNet) for multi-garment virtual dressing conditioned on any combination of garments and text prompts.
Problem
Existing garment-centric diffusion methods handle only single garments and struggle to combine multiple garments while keeping fine textures and text faithfulness.
Approach
Uses a Garment-Specific Feature Extractor that encodes each garment with its own LoRA-adapted self-attention inside a shared U-Net, a Dressing-Attention module with an instance-level localization loss routing each garment to its region, and a perceptual plus high-frequency texture loss.
Key results
- On multi-garment evaluation, AnyDressing reports CLIP-T 0.296, CLIP-I* 0.734, CLIP-AS 5.874 and DINO* 0.674, outperforming all baselines.
- User study with 40 participants preferred AnyDressing at 93.80% for texture consistency and 77.00% for prompt alignment.
- Ablation: GFE, IGL and GTL each improve CLIP-I*, CLIP-T or texture consistency; full model reaches CLIP-T 0.296 versus 0.260 for the base.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AnyDressing instead of the strongest baseline, StableGarment | Multi-garment virtual dressing · Dressing-Pair | CLIP-T: 0.284 -> 0.296 | Helped |
| Use AnyDressing instead of the strongest baseline, IMAGDressing | Multi-garment virtual dressing · Dressing-Pair | CLIP-I*: 0.614 -> 0.734 | Helped |
| Use AnyDressing instead of the strongest baseline, IP-Adapter | Multi-garment virtual dressing · Dressing-Pair | CLIP-AS: 5.795 -> 5.874 | Helped |
| Use AnyDressing instead of the strongest baseline, StableGarment | Multi-garment virtual dressing · Dressing-Pair | DINO*: 0.412 -> 0.674 | Helped |
| Add GFE to the base model (ablation) | AnyDressing base model · Dressing-Pair | CLIP-I*: 0.625 -> 0.718 | Helped |
| Add GFE to the base model (ablation) | AnyDressing base model · Dressing-Pair | CLIP-T: 0.260 -> 0.265 | No clear effect |
| Add IGL to Base+GFE (ablation) | AnyDressing · Dressing-Pair | CLIP-T: 0.265 -> 0.289 | Helped |
| Add IGL to Base+GFE (ablation) | AnyDressing · Dressing-Pair | CLIP-AS: 5.627 -> 5.790 | Helped |
| Add GTL to Base+GFE+IGL (ablation) | AnyDressing · Dressing-Pair | CLIP-I*: 0.722 -> 0.734 | Helped |
Limitations
- Baselines lack multi-garment support, so they were evaluated by concatenating garments spatially
- Training relies partly on a proprietary dataset that is not released
Experimental setup
Datasets: DressCode proprietary internet dataset VITON-HD
Benchmarks: VITON-HD garments (single-garment) proprietary single-garment set Dressing-Pair (multi-garment)
Metrics: CLIP-T CLIP-I CLIP-AS DINO CLIP-I* DINO* user preference
Baselines: IP-Adapter StableGarment MagicClothing IMAGDressing
Builds on: Stable Diffusion IP-Adapter DreamFit ReferenceNet LoRA
Closest papers by meaning
- DreamFit: Garment-Centric Human Generation via a Lightweight Anything-Dressing Encoder (AAAI 2025 · similarity 0.90)
- IMAGDressing-v1: Customizable Virtual Dressing (AAAI 2025 · similarity 0.90)
- Texture-Preserving Diffusion Models for High-Fidelity Virtual Try-On (CVPR 2024 · similarity 0.90)
- AnyFit: Controllable Virtual Try-on for Any Combination of Attire Across Any Scenario (NeurIPS 2024 · similarity 0.89)
- ITA-MDT: Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On (CVPR 2025 · similarity 0.89)
- Cascaded Diffusion Models for Virtual Try-On: Improving Control and Resolution (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.