Research map / Image generation / Editing and style transfer
Towards Squeezing-Averse Virtual Try-On via Sequential Deformation
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.
Identifies sleeve- and waist-squeezing artifacts in HR-VITON and proposes SD-VITON, which uses sequential appearance-flow deformation (TVOB + TACO layers), a masked tucked-out warping loss, and a 3D z-axis flow for erasing non-ROI clothing texture.
Problem
High-resolution image-based virtual try-on suffers from clothing texture squeezing at long sleeves and around the waist, caused by conflicting TV and adversarial losses in cascaded appearance flow estimation and by inability to erase clothing content.
Approach
Disentangles cascaded appearance flow prediction into TV-objective-dominant layers followed by a final task-coexistence layer with no residual path from earlier flows, and adds a 3D appearance flow with a zero-filled depth slice so that non-ROI texture can be erased without breaking flow smoothness.
Key results
- Achieves FID 9.75 and KID 0.120 on the VITON-HD test set, versus 10.43 FID for HR-VITON with composition mask (HR-VITON*).
- LPIPS improves from 0.0643 (HR-VITON*) to 0.0616 and SSIM from 0.8919 to 0.8955.
- Adds only 0.23M parameters over HR-VITON with identical inference speed (2.49 FPS on a single A6000).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a composition mask (ablation) | HR-VITON · VITON-HD test pairs | FID: 10.91 -> 10.43 | Helped |
| add a composition mask (ablation) | HR-VITON · VITON-HD test pairs | KID: 0.179 -> 0.179 | No clear effect |
| add sequential connection (ablation) | HR-VITON* · VITON-HD test pairs | FID: 10.43 -> 10.08; 0.35 points | Helped |
| add sequential connection (ablation) | HR-VITON* · long-sleeve test pairs, arm region | LPIPS of warped clothes | Helped |
| add non-ROI removal (ablation) | HR-VITON* + sequential connection · VITON-HD test pairs | FID: 10.08 -> 9.75 | Helped |
| add non-ROI removal (ablation) | HR-VITON* + sequential connection · tucked-in shirts test pairs, waist region | LPIPS of warped clothes | Helped |
| use SD-VITON instead of HR-VITON* | HR-VITON* · VITON-HD test pairs | FID: 10.43 -> 9.75; 0.68 points | Helped |
| use SD-VITON instead of HR-VITON* | HR-VITON* · VITON-HD test pairs | LPIPS: 0.0643 -> 0.0616 | Helped |
| use SD-VITON instead of HR-VITON* | HR-VITON* · VITON-HD test pairs | KID: 0.179 -> 0.120 | Helped |
| use SD-VITON instead of HR-VITON* | HR-VITON* · VITON-HD test pairs | SSIM: 0.8919 -> 0.8955 | Helped |
| use SD-VITON instead of HR-VITON | HR-VITON · inference on A6000 GPU | frames per second: 2.49 -> 2.49 | No clear effect |
| use SD-VITON instead of HR-VITON | HR-VITON · tucked-out shirts and short- or no-sleeve examples | visual quality | No clear effect |
Surprising findings
- The composition mask alone does not resolve the squeezing artifacts, though it improves FID from 10.91 to 10.43.
Limitations
- Evaluation limited to a single high-resolution dataset (VITON-HD) with mostly top clothing.
- Waist-squeezing fix relies on manually selected test subset (182 pairs) for body-part analysis.
- User study uses only 30 sets with 15 participants.
Experimental setup
Datasets: VITON-HD dataset
Benchmarks: VITON-HD test set (2,032 pairs)
Metrics: FID KID LPIPS SSIM user study selection ratio body-part LPIPS
Baselines: CP-VTON ACGPN PF-AFN VITON-HD HR-VITON
Builds on: HR-VITON VITON-HD PF-AFN
Closest papers by meaning
- Incorporating Visual Correspondence into Diffusion Model for Virtual Try-On (ICLR 2025 · similarity 0.88)
- InstructVVT: Instruction-Driven Video Virtual Try-On without Auxiliary Spatial Priors (NeurIPS 2026 · similarity 0.87)
- OOTDiffusion: Outfitting Fusion Based Latent Diffusion for Controllable Virtual Try-On (AAAI 2025 · similarity 0.87)
- M&M VTO: Multi-Garment Virtual Try-On and Editing (CVPR 2024 · similarity 0.87)
- Texture-Preserving Diffusion Models for High-Fidelity Virtual Try-On (CVPR 2024 · similarity 0.87)
- Cascaded Diffusion Models for Virtual Try-On: Improving Control and Resolution (AAAI 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.