Research map / Image generation / Editing and style transfer

Correpondence Alignment For Improved Virtual Try-On

NeurIPS 2026 · Image generation › Editing and style transfer · approach: 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 CORAL, a DiT-based virtual try-on method that distills DINOv3 person-to-garment correspondences into full 3D attention query-key matching and adds an attention entropy loss, improving SSIM/LPIPS/FID on VITON-HD and DressCode.

Problem

Virtual try-on struggles to transfer garment shape and fine details, especially in unpaired settings, because person-to-garment correspondence is learned only implicitly inside diffusion transformer attention.

Approach

Analyzes person-to-garment query-key attention in a FLUX-based DiT and explicitly supervises it with reliable cycle-consistent DINOv3 matches (soft-argmax L2 loss) plus an attention entropy minimization loss, rather than aligning intermediate features (REPA-style).

Key results

• On VITON-HD paired, CORAL reaches SSIM 0.907, LPIPS 0.048, FID 4.962, versus best baseline CatVTON FID 5.458 and LPIPS 0.058.
• On DressCode unpaired, CORAL achieves FID 4.692 vs IDM-VTON 5.343.
• Adding L_CORAL improves PCK by 34% while reducing attention entropy (per paper).

Surprising findings

• Directly aligning intermediate DiT features to DINOv3 (REPA) substantially degraded all metrics versus baseline, e.g., paired FID 13.718 vs 5.543.
• Attention entropy minimization alone improves perceptual metrics but does not fix where the model matches, so coarse garment properties remain wrong.

Limitations

• Relies on DINOv3 cycle-consistent pseudo ground-truth matches whose reliability depends on threshold choice.
• VLM-based evaluation depends on a single proprietary judge (GPT-5 mini).

Experimental setup

Datasets: VITON-HD DressCode PPR10K

Benchmarks: VITON-HD DressCode PPR10K

Metrics: SSIM LPIPS FID KID PCK GTC TAC FPC

Baselines: GPVTON StableVTON OOTDiffusion IDM-VTON CatVTON Any2AnyTryOn REPA

Builds on: FLUX.1-Fill-dev DINOv3 REPA rectified flow

Closest papers by meaning

• M&M VTO: Multi-Garment Virtual Try-On and Editing (CVPR 2024 · similarity 0.89)
• Incorporating Visual Correspondence into Diffusion Model for Virtual Try-On (ICLR 2025 · similarity 0.89)
• IMAGDressing-v1: Customizable Virtual Dressing (AAAI 2025 · similarity 0.88)
• MV-VTON: Multi-View Virtual Try-On with Diffusion Models (AAAI 2025 · similarity 0.88)
• Robust-MVTON: Learning Cross-Pose Feature Alignment and Fusion for Robust Multi-View Virtual Try-On (CVPR 2025 · similarity 0.87)
• CycleVTON: A Cycle Mapping Framework for Parser-Free Virtual Try-On (AAAI 2024 · similarity 0.87)