Research map / 3D vision / 3D generation
VTON 360: High-Fidelity Virtual Try-On from Any Viewing Direction
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 VTON 360, which extends a latent-diffusion 2D try-on model with multi-view garment and body inputs, SMPL-X normal-map poses, multi-view attention and camera-aware CLIP embeddings to produce 3D-consistent, any-view virtual try-on on Gaussian Splatting human models.
Problem
High-fidelity 3D virtual try-on that remains consistent and detailed when rendered from any viewing direction; existing 3D methods lose garment fidelity and 2D try-on models lack multi-view consistency.
Approach
Reformulate 3D VTON as multi-view 2D VTON: extend a 2D exemplar-based inpainting model with paired front/back garment inputs and multi-view body images, plus three techniques (SMPL-X normal-map pose, multi-view spatial attention weighted by camera angle, and camera-conditioned CLIP embedding), then reconstruct the edited views into a 3D Gaussian Splatting model.
Key results
- On Thuman2.0, VTON 360 reaches DINOsim 0.633 versus 0.577 for GaussCtrl and 0.556 for DreamWaltz, with CLIPcons 0.923 versus 0.931 for GaussCtrl.
- In a 50-participant user study, VTON 360 received 97.54% of quality votes and 96.46% of alignment votes on Thuman2.0.
- Ablation on Thuman2.0 shows CLIPcons rising from 0.892 (2D VTON baseline) to 0.923 and DINOsim from 0.609 to 0.633 as the three modules are added.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VTON 360 instead of the strongest baseline | 3D virtual try-on · Thuman2.0 | DINOsim: 0.577 -> 0.633 | Helped |
| Use VTON 360 instead of the strongest baseline | 3D virtual try-on · MVHumanNet | DINOsim: 0.521 -> 0.623 | Helped |
| Use VTON 360 instead of the strongest baseline | 3D virtual try-on · Thuman2.0 | Votequality: 1.08% -> 97.54% | Helped |
| Use VTON 360 instead of the strongest baseline | 3D virtual try-on · MVHumanNet | Votequality: 2.15% -> 95.69% | Helped |
| Use VTON 360 instead of the strongest baseline | 3D virtual try-on · Thuman2.0 | CLIPcons: 0.939 -> 0.923 | No clear effect |
| Use VTON 360 instead of the strongest baseline | 3D virtual try-on · MVHumanNet | CLIPcons: 0.948 -> 0.933 | No clear effect |
| Add pseudo-3D pose (ablation) | 2D-VTON · Thuman2.0 | DINOsim: 0.609 -> 0.626 | Helped |
| Add multi-view CLIP embedding (ablation) | 2D-VTON + Pseudo-3D Pose · Thuman2.0 | DINOsim: 0.626 -> 0.631 | Helped |
| Add multi-view spatial attention (ablation) | 2D-VTON + Pseudo-3D Pose + Multi-view CLIP Embedding · Thuman2.0 | DINOsim: 0.631 -> 0.633 | Helped |
Surprising findings
- Baselines with SDS loss get higher CLIP directional consistency than VTON 360, which the authors attribute to SDS smoothing and flatter textures rather than better consistency.
Experimental setup
Datasets: THuman2.0 MVHumanNet MVG
Benchmarks: THuman2.0 MVHumanNet
Metrics: DINO similarity CLIP directional consistency score user study preference vote
Baselines: DreamWaltz GaussCtrl TIP-Editor 2D VTON baseline (Yang et al.)
Builds on: Stable Diffusion 3D Gaussian Splatting SMPL-X GaussianVTON DreamVTON
Closest papers by meaning
- MV-VTON: Multi-View Virtual Try-On with Diffusion Models (AAAI 2025 · similarity 0.91)
- OpenVTON-Bench: A Large-Scale High-Resolution Benchmark for Controllable Virtual Try-On Evaluation (NeurIPS 2026 · similarity 0.88)
- Incorporating Visual Correspondence into Diffusion Model for Virtual Try-On (ICLR 2025 · similarity 0.88)
- BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training (CVPR 2025 · similarity 0.88)
- IPVTON: Image-based 3D Virtual Try-on with Image Prompt Adapter (AAAI 2025 · similarity 0.87)
- M&M VTO: Multi-Garment Virtual Try-On and Editing (CVPR 2024 · 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.