Research map / 3D vision / 3D generation
Robust-MVTON: Learning Cross-Pose Feature Alignment and Fusion for Robust Multi-View Virtual Try-On
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 Robust-MVTON, an end-to-end multi-view virtual try-on method that fuses front- and back-view clothing features via learned cross-pose projectors with pose and spatial-label losses, feeding a coarse-to-fine latent diffusion model.
Problem
Multi-view virtual try-on: generating try-on images for arbitrary target views (including side and back) from front- and back-view garment images, where naive concatenation or warping-based fusion fails.
Approach
A cross-pose feature alignment module uses two linear projectors (one per input view) that map front/back pose, clothing, and spatial-label features into the target view space, supervised by a pose alignment loss and a label loss, so the fused clothing features are adaptively aligned before a coarse-to-fine latent diffusion model generates the result.
Key results
- On DeepFashion, Robust-MVTON reaches FID 11.40, LPIPS 0.158, SSIM 0.820, beating CFLD* (FID 14.31) and MV-VTON (FID 21.13).
- On MPV, Robust-MVTON achieves FID 12.51 and SSIM 0.818 versus CFLD* FID 14.36 and MV-VTON FID 19.20.
- User study preference: 75.2% on DeepFashion and 83.3% on MPV for Robust-MVTON.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · DeepFashion | FID: 14.31 -> 11.40 | Helped |
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · DeepFashion | LPIPS: 0.181 -> 0.158 | Helped |
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · DeepFashion | SSIM: 0.782 -> 0.820 | Helped |
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · DeepFashion | User Study: 12.6% -> 75.2% | Helped |
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · MPV | FID: 14.36 -> 12.51 | Helped |
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · MPV | LPIPS: 0.185 -> 0.174 | Helped |
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · MPV | SSIM: 0.779 -> 0.818 | Helped |
| use Robust-MVTON instead of CFLD* | Multi-view virtual try-on · MPV | User Study: 6.5% -> 83.3% | Helped |
| add spatial label assistance (ablation) | Robust-MVTON · DeepFashion | FID: 11.42 -> 11.40 | Helped |
| add pose loss constraint (ablation) | Robust-MVTON · DeepFashion | FID: 12.45 -> 11.40 | Helped |
Surprising findings
- Directly concatenating front and back views into CFLD (CFLD*) is not an effective way to embed multi-view reference features, and MV-VTON's warping-based pipeline performed worse than even frontal-only baselines.
Limitations
- Limited by scarcity of suitable 2D multi-view try-on datasets; MPV mostly has only front and back pairs.
- Relies on the coarse-to-fine CFLD architecture and external parsing/pose tools.
- Evaluated only at 512x384 resolution on two datasets.
Experimental setup
Datasets: DeepFashion MPV
Benchmarks: DeepFashion MPV
Metrics: FID LPIPS SSIM user study
Baselines: CAT-DM StableVITON MV-VTON CFLD*
Builds on: CFLD Stable Diffusion MV-VTON OpenPose Swin Transformer
Closest papers by meaning
- MV-VTON: Multi-View Virtual Try-On with Diffusion Models (AAAI 2025 · similarity 0.93)
- Pursuing Temporal-Consistent Video Virtual Try-On via Dynamic Pose Interaction (CVPR 2025 · similarity 0.88)
- IPVTON: Image-based 3D Virtual Try-on with Image Prompt Adapter (AAAI 2025 · similarity 0.88)
- Correpondence Alignment For Improved Virtual Try-On (NeurIPS 2026 · similarity 0.87)
- M&M VTO: Multi-Garment Virtual Try-On and Editing (CVPR 2024 · similarity 0.87)
- Incorporating Visual Correspondence into Diffusion Model for Virtual Try-On (ICLR 2025 · similarity 0.87)
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.