Research map / Image generation / Editing and style transfer
Latent Diffusion-Enhanced Virtual Try-On via Optimized Pseudo-Label Generation
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 a latent-diffusion virtual try-on framework that generates pseudo-person labels via an inpainting pretext task, refines them with a noise-covering strategy, and trains a supervised try-on synthesizer with a skin refinement loss.
Problem
Virtual try-on lacks paired ground-truth training data, so existing self-supervised inpainting methods depend on error-prone clothing-agnostic masks and parsers, which degrade try-on quality.
Approach
Use a pretext inpainting LDM to produce pseudo-person images as paired supervision, then cover the unreliable clothing region with the warped target garment (noise covering) and add a VGG-based skin refinement loss so a downstream LDM can be trained fully supervised with the real person as ground truth.
Key results
- On VITON-HD, the method reports SSIM 0.898, LPIPS 0.061, FID 9.313, KID 1.21, the best among compared methods including StableVITON (FID 9.465) and DCI-VTON (FID 8.754 but worse SSIM/KID).
- On VITON, FID 9.86 and SSIM 0.91, the best FID among listed baselines (USC-PFN 10.47).
- Ablation on VITON-HD: removing clothing pre-alignment raises FID from 9.313 to 10.984 and KID from 1.21 to 2.29.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Ours instead of USC-PFN | VITON virtual try-on · VITON | SSIM: 0.91 -> 0.91 | No clear effect |
| use Ours instead of PF-AFN | VITON virtual try-on · VITON | FID: 10.09 -> 9.86 | Helped |
| use Ours instead of KGI | VITON-HD virtual try-on · VITON-HD | SSIM: 0.892 -> 0.898 | Helped |
| use Ours instead of DCI-VTON | VITON-HD virtual try-on · VITON-HD | FID: 8.754 -> 9.313 | Hurt |
| use Ours instead of DCI-VTON | VITON-HD virtual try-on · VITON-HD | KID: 1.10 -> 1.21 | Hurt |
| add clothing pre-alignment (ablation) | Ours · VITON-HD | FID: 10.984 -> 9.313 | Helped |
| add noise covering (ablation) | Ours · VITON-HD | FID: 10.097 -> 9.313 | Helped |
| add skin refinement loss (ablation) | Ours · VITON-HD | FID: 9.427 -> 9.313 | Helped |
| add clothing pre-alignment (ablation) | Ours · VITON-HD | KID: 2.29 -> 1.21 | Helped |
| add noise covering (ablation) | Ours · VITON-HD | KID: 1.88 -> 1.21 | Helped |
| add skin refinement loss (ablation) | Ours · VITON-HD | KID: 1.43 -> 1.21 | Helped |
Surprising findings
- DCI-VTON achieves a lower FID (8.754) than the proposed method on VITON-HD, though the proposed method leads on SSIM, LPIPS and KID.
Limitations
- Performance limited by scarcity and diversity of training clothing samples.
- Performance limited by variability and unpredictability of human poses.
- Pseudo-labels from the pretext task can still be coarse and contain errors outside the clothing region.
Experimental setup
Datasets: VITON-HD VITON
Benchmarks: VITON-HD VITON
Metrics: SSIM LPIPS FID KID
Baselines: ACGPN LM-VTON DCTON PF-AFN ZFlow RT-VTON SDAFN DressCode PL-VTON POVNet USC-PFN VITON-HD
Builds on: Stable Diffusion PbE Classifier-free guidance LaDI-VTON
Closest papers by meaning
- BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training (CVPR 2025 · similarity 0.87)
- Texture-Preserving Diffusion Models for High-Fidelity Virtual Try-On (CVPR 2024 · similarity 0.87)
- OOTDiffusion: Outfitting Fusion Based Latent Diffusion for Controllable Virtual Try-On (AAAI 2025 · similarity 0.87)
- Cascaded Diffusion Models for Virtual Try-On: Improving Control and Resolution (AAAI 2025 · similarity 0.87)
- CycleVTON: A Cycle Mapping Framework for Parser-Free Virtual Try-On (AAAI 2024 · similarity 0.87)
- CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models (ICLR 2025 · similarity 0.86)
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.