Research map / 3D vision / 3D generation
MV-VTON: Multi-View Virtual Try-On with 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 MV-VTON, a diffusion-based virtual try-on method that uses frontal and back garment images with view-adaptive selection and joint attention blocks, plus a new multi-view MVG dataset.
Problem
Existing image-based virtual try-on only handles frontal person views with frontal clothing, failing when the person's view is non-frontal and clothing views are inconsistent.
Approach
Uses both frontal and back garment images; view-adaptive selection (hard selection for global CLIP features, soft selection for multi-scale local features) chooses clothing views according to pose similarity, and joint attention blocks align and selectively fuse global and local clothing features in a Paint-by-Example-based inpainting diffusion model.
Key results
- On the MVG multi-view dataset (paired setting), MV-VTON achieves LPIPS 0.050, SSIM 0.936, FID 22.18 versus best baseline StableVITON LPIPS 0.063, SSIM 0.929, FID 23.52.
- On VITON-HD paired setting, MV-VTON reports LPIPS 0.069, SSIM 0.897, FID 5.43, KID 0.49, improving over DCI-VTON (FID 5.52).
- Unpaired FID_u on MVG drops to 33.44 from 35.85 for the best prior method (StableVITON).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MV-VTON instead of the strongest reported baseline | MV-VTON · MVG paired | LPIPS: 0.063 -> 0.050 | Helped |
| Use MV-VTON instead of the strongest reported baseline | MV-VTON · VITON-HD paired | LPIPS: 0.073 -> 0.069 | Helped |
| Use MV-VTON instead of the strongest reported baseline | MV-VTON · DressCode - Upper Body paired | SSIM: 0.937 -> 0.941 | Helped |
| Use MV-VTON instead of the strongest reported baseline | MV-VTON · MVG unpaired | FID_u: 35.85 -> 33.44 | Helped |
| Add soft-selection to hard-selection (ablation) | MV-VTON · MVG | LPIPS: 0.052 -> 0.050 | Helped |
| Add hard-selection to soft-selection (ablation) | MV-VTON · MVG | LPIPS: 0.064 -> 0.050 | Helped |
| Add hard-selection and soft-selection (ablation) | MV-VTON · MVG | LPIPS: 0.068 -> 0.050 | Helped |
| Add the local feature branch to global features (ablation) | MV-VTON joint attention · MVG | LPIPS: 0.062 -> 0.050 | Helped |
| Add the global feature branch to local features (ablation) | MV-VTON joint attention · MVG | LPIPS: 0.058 -> 0.050 | Helped |
| Use both global and local feature branches instead of global features only (ablation) | MV-VTON joint attention · VITON-HD | LPIPS: 0.074 -> 0.069 | Helped |
Surprising findings
- On the DressCode upper-body paired setting, StableVITON achieves a slightly better LPIPS (0.040 vs 0.040 tie) and the paper's own method only ties it on that metric, indicating limited gains on that single metric.
Limitations
- Struggles to preserve small or complex garment details, which are lost during latent-space inpainting.
- Requires both frontal and back garment images as input for multi-view try-on.
- Baselines on MVG were fine-tuned by the authors, and MVG is built from only 1,009 videos.
Experimental setup
Datasets: MVG VITON-HD DressCode
Benchmarks: MVG VITON-HD DressCode
Metrics: LPIPS SSIM FID KID FID_u KID_u
Baselines: Paint by Example PF-AFN GP-VTON LaDI-VTON DCI-VTON StableVITON IDM-VTON
Builds on: Paint by Example Stable Diffusion StableVITON CLIP
Closest papers by meaning
- Robust-MVTON: Learning Cross-Pose Feature Alignment and Fusion for Robust Multi-View Virtual Try-On (CVPR 2025 · similarity 0.93)
- VTON 360: High-Fidelity Virtual Try-On from Any Viewing Direction (CVPR 2025 · similarity 0.91)
- Incorporating Visual Correspondence into Diffusion Model for Virtual Try-On (ICLR 2025 · similarity 0.91)
- M&M VTO: Multi-Garment Virtual Try-On and Editing (CVPR 2024 · similarity 0.90)
- BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training (CVPR 2025 · similarity 0.90)
- IMAGDressing-v1: Customizable Virtual Dressing (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.