Research map / Image generation / Editing and style transfer
CycleVTON: A Cycle Mapping Framework for Parser-Free 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 CycleVTON, a parser-free shared-weight cycle-mapping framework for image-based virtual try-on that uses a flow-constraint loss and skin generation strategy to enable supervised-style training on arbitrarily matched clothing and person images.
Problem
Image-based virtual try-on lacks ground truth for arbitrarily matched clothing and person pairs, and existing cycle-consistency and parser-based pipelines are hard to converge and are harmed by flawed human parsing.
Approach
A single shared-weight cycle mapping network (forward and reverse pipelines) with a flow-constraint loss that supervises the deformation field using a Res-UNet trained on DensePose and paired clothing, plus a skin generation strategy that removes and refills skin regions, and no human parser input.
Key results
- On VITON unpaired setting, FID 19.64 and KID 0.37 for clothing deformation, beating the best AF-based baseline (Flow-Style) by 0.74 FID.
- Try-on synthesis SSIM of 0.92 (paired) and FID of 9.41 on VITON, improving on SDAFN (0.88 SSIM, 12.05 FID in the table) and RT-VTON.
- Computational cost: 87M params and 29G FLOPs on VITON at 39 FPS, about half the FLOPs of PF-AFN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CycleVTON instead of Flow-Style for clothing deformation | clothing deformation · VITON, paired | FID: 20.07 -> 18.44 | Helped |
| use CycleVTON instead of Flow-Style for clothing deformation | clothing deformation · VITON, unpaired | FID: 20.38 -> 19.64 | Helped |
| use CycleVTON instead of HR-VITON for clothing deformation | clothing deformation · VITON-HD, paired | FID: 23.51 -> 19.15 | Helped |
| use CycleVTON instead of HR-VITON for clothing deformation | clothing deformation · VITON-HD, unpaired | FID: 24.83 -> 22.51 | Helped |
| use CycleVTON instead of PF-AFN for paired try-on | try-on synthesis · VITON, paired | SSIM: 0.89 -> 0.92 | Helped |
| use CycleVTON instead of PF-AFN for unpaired try-on | try-on synthesis · VITON, unpaired | FID: 10.21 -> 9.41 | Helped |
| add flow-constraint loss to the forward pipeline (ablation) | CycleVTON · VITON clothing deformation | clothing deformation quality | Helped |
| add the skin generation strategy (ablation) | CycleVTON · VITON try-on synthesis | skin generation | Helped |
| remove flow-constraint loss and skin generation strategy (ablation) | CycleVTON · VITON try-on synthesis | try-on quality | Hurt |
Surprising findings
- The paper reports that the framework without the flow constraint loss and skin generation strategy does not work, per the ablation figure.
Limitations
- Relies on a pretrained semantic map generator for dynamic data interception, which reintroduces a dependence on segmentation.
- Pipeline depends on pseudo ground-truth flow from a pretrained Res-UNet (omega), limiting deformation beyond its prior.
- Evaluated only at 256x192 on VITON and VITON-HD; high-resolution performance is reported only via downsampling.
Experimental setup
Datasets: VITON VITON-HD
Benchmarks: VITON VITON-HD
Metrics: FID KID SSIM FPS FLOPs #Params
Baselines: VITON CP-VTON ACGPN DCTON PF-AFN Flow-Style HR-VITON VITON-HD SDAFN RT-VTON
Builds on: CycleGAN USC-PFN DCTON Res-UNet appearance flow
Closest papers by meaning
- Enhancing Virtual Try-On with Synthetic Pairs and Error-Aware Noise Scheduling (CVPR 2025 · similarity 0.89)
- IPVTON: Image-based 3D Virtual Try-on with Image Prompt Adapter (AAAI 2025 · similarity 0.88)
- Virtual Try-On: Real-Time Interactive Hybrid Network with High-Fidelity (AAAI 2024 · similarity 0.88)
- MV-VTON: Multi-View Virtual Try-On with Diffusion Models (AAAI 2025 · similarity 0.88)
- BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training (CVPR 2025 · similarity 0.88)
- IMAGDressing-v1: Customizable Virtual Dressing (AAAI 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.