Research map / Image generation / Editing and style transfer
OOTDiffusion: Outfitting Fusion Based Latent Diffusion for Controllable 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 OOTDiffusion, a latent-diffusion virtual try-on model that uses a parallel outfitting UNet and self-attention fusion (no warping) plus outfitting dropout for garment-strength classifier-free guidance.
Problem
Image-based virtual try-on needs realistic outfitted images that preserve fine garment details (text, patterns) without relying on an explicit warping step that overfits training data.
Approach
A duplicated pretrained SD UNet (outfitting UNet) encodes the garment latent in one step; its self-attention features are concatenated spatially with the denoising UNet's features (outfitting fusion), and outfitting dropout (10% zeroed garment latents) enables classifier-free guidance on garment strength.
Key results
- On VITON-HD (512x384), OOTDiffusion reaches LPIPS 0.071, FID 8.81, KID 0.82, beating GP-VTON (LPIPS 0.083, FID 9.17) and StableVITON (FID 9.13).
- On Dress Code (all categories), OOTDiffusion achieves FID 4.20 and KID 0.37, versus LaDI-VTON FID 5.66 and GP-VTON FID 5.88.
- In cross-dataset Dress Code to VITON-HD evaluation, FID drops to 11.22 vs. 29.66 for LaDI-VTON.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use OOTDiffusion instead of GP-VTON | Virtual try-on · VITON-HD | LPIPS: 0.083 -> 0.071 | Helped |
| use OOTDiffusion instead of GP-VTON | Virtual try-on · VITON-HD | SSIM: 0.892 -> 0.878 | Hurt |
| use OOTDiffusion instead of StableVITON | Virtual try-on · VITON-HD | FID: 9.13 -> 8.81 | Helped |
| use OOTDiffusion instead of GP-VTON | Virtual try-on · VITON-HD | KID: 0.93 -> 0.82 | Helped |
| use OOTDiffusion instead of GP-VTON | Virtual try-on · Dress Code (All) | LPIPS: 0.051 -> 0.045 | Helped |
| use OOTDiffusion instead of GP-VTON | Virtual try-on · Dress Code (All) | SSIM: 0.921 -> 0.927 | Helped |
| use OOTDiffusion instead of LaDI-VTON | Virtual try-on · Dress Code (All) | FID: 5.66 -> 4.20 | Helped |
| use OOTDiffusion instead of LaDI-VTON | Virtual try-on · Dress Code (All) | KID: 1.21 -> 0.37 | Helped |
| add outfitting dropout and use guidance scale 1.5 instead of no dropout (ablation) | OOTDiffusion · VITON-HD | LPIPS: 0.0750 -> 0.0705 | Helped |
| add outfitting dropout and use guidance scale 1.5 instead of no dropout (ablation) | OOTDiffusion · VITON-HD | SSIM: 0.8699 -> 0.8775 | Helped |
| add outfitting dropout and use guidance scale 1.5 instead of no dropout (ablation) | OOTDiffusion · VITON-HD | FID: 8.91 -> 8.81 | Helped |
| add outfitting dropout and use guidance scale 1.5 instead of no dropout (ablation) | OOTDiffusion · VITON-HD | KID: 0.89 -> 0.82 | Helped |
Surprising findings
- Guidance scales above about 2.5 cause color distortion, with severe distortion at 5.0.
- GP-VTON falls far behind other methods in cross-dataset evaluation, which the authors attribute to its warping module overfitting.
Limitations
- Trained on paired human and garment images, so cross-category try-on (e.g., T-shirt on a long dress) may fail.
- Original body details such as muscles, watches or tattoos may be altered because the masked region is repainted.
Experimental setup
Datasets: VITON-HD Dress Code
Benchmarks: VITON-HD Dress Code
Metrics: LPIPS SSIM FID KID
Baselines: VITON-HD HR-VITON GP-VTON LaDI-VTON StableVITON Paint-by-Example PSAD
Builds on: Stable Diffusion classifier-free guidance CLIP textual inversion UniPC
Closest papers by meaning
- Incorporating Visual Correspondence into Diffusion Model for Virtual Try-On (ICLR 2025 · similarity 0.90)
- IMAGDressing-v1: Customizable Virtual Dressing (AAAI 2025 · similarity 0.89)
- Cascaded Diffusion Models for Virtual Try-On: Improving Control and Resolution (AAAI 2025 · similarity 0.89)
- M&M VTO: Multi-Garment Virtual Try-On and Editing (CVPR 2024 · similarity 0.89)
- Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling (CVPR 2025 · similarity 0.88)
- BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training (CVPR 2025 · similarity 0.88)
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.