CORAL: Correspondence Alignment for Improved Virtual Try-On

Jiyoung Kim; Youngjin Shin; Siyoon Jin; Dahyun Chung; Jisu Nam; Tongmin Kim; Jongjae Park; Hyeonwoo Kang

cs.CVFeb 19, 2026

CORAL: Correspondence Alignment for Improved Virtual Try-On

Authors: Jiyoung Kim, Youngjin Shin, Siyoon Jin, Dahyun Chung, Jisu Nam, Tongmin Kim, Jongjae Park, Hyeonwoo Kang,

Organizations: KAIST AI · NC AI Co., Ltd.

Abstract

Existing methods for Virtual Try-On (VTON) often struggle to preserve fine garment details, especially in unpaired settings where accurate person-garment correspondence is required. These methods do not explicitly enforce person-garment alignment and fail to explain how correspondence emerges within Diffusion Transformers (DiTs). In this paper, we first analyze full 3D attention in DiT-based architecture and reveal that the person-garment correspondence critically depends on precise person-garment query-key matching within the full 3D attention. Building on this insight, we then introduce CORrespondence ALignment (CORAL), a DiT-based framework that explicitly aligns query-key matching with robust external correspondences. CORAL integrates two complementary components: a correspondence distillation loss that aligns reliable matches with person-garment attention, and an entropy minimization loss that sharpens the attention distribution. We further propose a VLM-based evaluation protocol to better reflect human preference. CORAL consistently improves over the baseline, enhancing both global shape transfer and local detail preservation. Extensive ablations validate our design choices.

Figures & tables

Figure 2: Correlation between Query-Key Matching and VTON Performance. Pink marker denotes the query points. (a) presents qualitative correlation between VTON performance and person $\to$ garment attention. All outputs in (a) are generated by the baseline. Human-preferred outputs show accurately localized, sharp attention, while low quality outputs show dispersed attention on incorrect locations. (b) reports quantitative correlation, where $r$ denotes Pearson correlation coefficient. PCK of person $\to$ garment correspondence correlates with SSIM and LPIPS of the generated images, indicating that better localization predicts higher visual quality (Best viewed when zoomed-in).

Figure 3: Overall Architecture. CORAL builds upon a baseline architecture that constructs the noisy latent $\mathbf{z}_{t}$ by horizontally concatenating the noisy garment latents $\mathbf{z}_{\text{g},t}$ and person latents $\mathbf{z}_{\text{p},t}$ , and then channel-wise concatenates the conditioning canvas $\mathbf{z}_{\text{diptych}}$ and mask canvas $\mathbf{m}_{\text{diptych}}$ with $\mathbf{z}_{t}$ before the input projection layer. Pose is injected by adding $\mathbf{z}_{\text{pose}}$ as tokens, with RoPE set to share spatial positions between person and pose tokens. $\mathcal{L}_{\text{CORAL}}$ is applied to the person $\rightarrow$ garment matching cost $A^{t,l}_{\mathcal{P}\rightarrow\mathcal{G}}$ estimated from MM-Attention within DiT blocks: $\mathcal{L}_{\text{corr}}$ aligns $A^{t,l}_{\mathcal{P}\rightarrow\mathcal{G}}$ to pseudo ground-truth correspondences extracted from DINOv3, while $\mathcal{L}_{\text{ent}}$ is computed on $A^{t,l}_{\mathcal{P}\rightarrow\mathcal{G}}$ to encourage sharper, more localized matches.

Figure 4: Correspondence Visualization and Warped Results. We visualize correspondence fields and the resulting warped garment, computed by mapping pixels from the garment reference image to the garment region of the person image. (a) Baseline attention-derived correspondence. (b) DINOv3 correspondence before reliability filtering. (c) Refined correspondence after cycle-consistency check. The baseline warp shows geometric distortion, while unfiltered DINOv3 can mistakenly match visually similar regions, such as lower-garment areas.

Figure 5: Qualitative Comparison. We show qualitative results on the standard benchmarks VITON-HD ( Choi et al., 2021 ) and DressCode ( Morelli et al., 2022 ) , as well as in-the-wild evaluation dataset built from PPR10K ( Liang et al., 2021 ) (Best viewed when zoomed-in). Additional qualitative results are provided in Appendix J

Methods	Paired				Unpaired
Methods	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
GPVTON ( Xie et al., 2023 )	0.878	0.067	8.938	4.257	11.993	4.570
StableVTION ( Kim et al., 2023 )	0.888	0.073	8.233	0.490	9.026	3.029
OOTDiffusion ( Kim et al., 2023 )	0.842	0.087	6.619	0.845	9.938	1.302
IDM-VTON ( Choi et al., 2024 )	0.866	0.062	6.009	0.838	9.198	1.203
CatVTON ( Chong et al., 2025 )	0.874	0.058	5.458	0.439	9.076	1.184

Table 1: Quantitative Comparison on VITON-HD ( Choi et al., 2021 ) .

Methods	Paired				Unpaired
Methods	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
GPVTON ( Xie et al., 2023 )	0.918	0.068	8.423	2.439	9.144	3.936
OOTDiffusion ( Xu et al., 2024 )	0.886	0.069	5.082	1.377	9.276	4.009
IDM-VTON ( Choi et al., 2024 )	0.904	0.052	3.472	0.882	5.343	1.321
CatVTON ( Chong et al., 2025 )	0.875	0.075	5.384	1.903	7.998	3.242
Any2AnyTryOn ∗ ( Guo et al., 2025 )	-	-	-	-	5.573	1.458

Table 2: Quantitative Comparison on DressCode ( Morelli et al., 2022 ) . ∗ does not support paired setting on specific dataset.

Methods	Paired				Unpaired
Methods	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
OOTDiffusion ( Kim et al., 2023 )	0.804	0.130	99.492	15.592	87.818	15.467
IDM-VTON ( Choi et al., 2024 )	0.844	0.111	64.250	1.911	63.638	3.600
CatVTON ( Chong et al., 2025 )	0.743	0.147	81.722	10.979	76.417	11.207
Any2AnyTryOn ∗ ( Guo et al., 2025 )	-	-	-	-	49.728	0.421
CORAL (w/o $\mathcal{L}_{\text{CORAL}}$ )	0.877	0.078	44.117	0.015	56.644	1.202

Table 3: Quantitative Comparison on PPR10K ( Liang et al., 2021 ) . ∗ does not support paired setting on specific dataset.

Train/Test	VITON-HD/VITON-HD			DressCode/DressCode			DressCode/PPR10K
Train/Test	GTC $\uparrow$	TAC $\uparrow$	FPC $\uparrow$	GTC $\uparrow$	TAC $\uparrow$	FPC $\uparrow$	GTC $\uparrow$	TAC $\uparrow$	FPC $\uparrow$
GPVTON ∗ ( Xie et al., 2023 )	3.93	4.28	4.17	3.19	3.31	3.14	-	-	-
OOTDiffusion ( Xu et al., 2024 )	3.97	4.34	4.12	3.25	3.24	3.45	1.04	2.06	1.57
IDM-VTON ( Choi et al., 2024 )	3.96	4.40	4.25	3.25	3.28	3.61	1.72	2.21	2.37
CatVTON ( Chong et al., 2025 )	3.82	4.20	4.06	3.07	3.22	3.53	1.05	2.00	2.01
Any2AnyTryOn ( Guo et al., 2025 )	3.81	4.27	4.19	2.99	3.28	3.36	1.92	2.45	2.48

Table 4: VLM-based evaluation on unpaired settings. ∗ does not support inference on specific dataset.

Figure 6: Ablation of Loss Components. We demonstrate the effectiveness of the two losses, $\mathcal{L}_{\text{ent}}$ and $\mathcal{L}_{\text{corr}}$ . Orange and green markers denote query points, and attention maps outlined in the same colors indicate the matches for each variant. By combining $\mathcal{L}_{\text{ent}}$ with $\mathcal{L}_{\text{corr}}$ , our model localizes the correct keys most accurately and exhibits the sharpest attention, yielding the best VTON performance.

Figure 7: Effect of $\mathcal{L}_{\text{CORAL}}$ . We visualize person $\rightarrow$ garment correspondences from DiT attention and the resulting try-on images before and after applying $\mathcal{L}_{\text{CORAL}}$ . The rightmost plot shows correspondence quality measured by PCK ( $\alpha=16$ ) and attention entropy, averaged across timesteps and layers.

	$\mathcal{L}_{\text{corr}}$	$\mathcal{L}_{\text{ent}}$	Paired				Unpaired
	$\mathcal{L}_{\text{corr}}$	$\mathcal{L}_{\text{ent}}$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
(I)	✗	✗	0.889	0.055	5.543	0.870	9.641	1.423
(II)	✗	✓	0.905	0.052	5.482	0.986	9.012	1.011
(III)	✓	✗	0.906	0.051	5.451	1.028	8.979	1.250
(IV)	✓	✓	0.907	0.048	4.962	0.565	8.763	0.880

Table 5: Ablation of Loss Components on VITON-HD.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	$\mathcal{L}_{\text{CORAL}}$	Paired				Unpaired
Methods	$\mathcal{L}_{\text{CORAL}}$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
Pose Token Concat	✗	0.889	0.055	5.543	0.870	9.641	1.323
Pose Token Concat	✓	0.907 ( +2.02% )	0.048 ( +12.73% )	4.962 ( +10.48% )	0.565 ( +35.06% )	8.763 ( +9.11% )	0.880 ( +33.48% )
Pose Channel Concat	✗	0.886	0.060	5.554	0.540	9.841	1.221
Pose Channel Concat	✓	0.888 ( +0.23% )	0.059 ( +1.67% )	5.372 ( +3.28% )	0.744	9.323 ( +5.26% )	1.077 ( +11.79% )

Appendix

Table C.1 : Effect of

\mathcal{L}_{\text{CORAL}}

across Baselines with Different Pose Injection Strategies for VITON-HD ( Choi et al., 2021 ) .

Figure C.2 : Effect of $\mathcal{L}_{\text{CORAL}}$ across Baselines with Different Pose Injection Strategies.

Figure C.3 : Efficient Baseline Architecture.

\mathcal{L}_{\text{CORAL}}

VITON-HD

DressCode

Paired

Unpaired

Paired

Unpaired

SSIM

\uparrow

LPIPS

\downarrow

FID

\downarrow

KID

\downarrow

FID

\downarrow

KID

\downarrow

SSIM

\uparrow

LPIPS

\downarrow

FID

\downarrow

KID

\downarrow

FID

\downarrow

KID

\downarrow

✗

0.864

0.084

9.521

1.353

11.251

0.456

0.905

0.065

5.752

0.473

8.723

0.736

✓

0.871 ( +0.81% )

0.084

6.271 ( +34.1% )

0.563 ( +58.5% )

10.285 ( +8.62% )

0.315 ( +30.9% )

0.905

0.061 ( +6.15% )

5.221 ( +9.22% )

0.581

8.70 ( +0.23% )

0.193 ( +73.78% )

Appendix

Table C.2 : Effect of

\mathcal{L}_{\text{CORAL}}

on Additional Efficient Baseline for VITON-HD ( Choi et al., 2021 ) and DressCode ( Morelli et al., 2022 ) .

Setting	$\mathcal{L}_{\text{CORAL}}$	GPU	GPU# (T/I)	Batch (T/I)	Speed (T/I)	VRAM (T/I)
Token-concat	✓	B200	8 / 1	16 / 1	2.61 / 0.091	148 / 40
Token-concat	✗	B200	8 / 1	16 / 1	3.01 / 0.105	148 / 40
Channel-concat	✓	B200	8 / 1	16 / 1	3.41 / 0.119	148 / 40
Channel-concat	✗	B200	8 / 1	16 / 1	3.92 / 0.137	148 / 40
Efficient	✓	A100	1 / 1	1 / 1	0.16 / 0.025	48 / 33
Efficient	✗	A100	1 / 1	1 / 1	0.18 / 0.025	48 / 33

Appendix

Table C.3 : Speed and memory. T denotes training and I denotes inference. Speed is in img/sec and VRAM is in GB.

	Methods	Paired				Unpaired
	Methods	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
(I)	Baseline	0.889	0.055	5.543	0.870	9.641	1.323
(II)	(I) + Feature Alignment (REPA) ( Yu et al., 2025 )	0.885 ( ${-0.45\%}$ )	0.087 ( ${-58.18\%}$ )	13.718 ( ${-147.48\%}$ )	2.103 ( ${-141.72\%}$ )	15.621 ( ${-62.03\%}$ )	3.684 ( ${-178.46\%}$ )
(III)	(I) + Query-Key Correspondence Alignment (Ours)	0.907 ( +2.02% )	0.048 ( +12.73% )	4.962 ( +10.48% )	0.565 ( +35.06% )	8.763 ( +9.11% )	0.880 ( +33.48% )

Appendix

Table E.4 : Alignment Target Ablation on VITON-HD ( Choi et al., 2021 ) .

Figure E.4 : Intermediate Feature Alignment vs. Ours.

Figure F.5 : VLM-based Evaluation Results Comparison.

Figure G.6 : User Study Results.

Figure H.7 : Additional Evaluation Dataset Samples.

Figure I.8 : Person-To-Person Garment Transfer Motivation.

Figure I.9 : Person-To-Person Architecture.

Methods	Paired				Unpaired
Methods	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
CORAL (w/o $\mathcal{L}_{\text{CORAL}}$ )	0.877	0.078	44.117	0.015	56.644	1.202
CORAL (w $\mathcal{L}_{\text{CORAL}})$	0.915	0.060	43.648	0.011	53.164	1.101
CORAL-P2P (w $\mathcal{L}_{\text{CORAL}})$	0.923	0.050	38.590	0.008	52.047	1.1088

Appendix

Table I.5 : Quantitative Comparison on PPR10K.

Figure I.10 : Qualitative Results of CORAL-P2P on PPR10K.

Figure J.11 : Additional Qualitative Comparison of CORAL on VITON-HD ( Choi et al., 2021 ) .

Figure J.12 : Additional Qualitative Comparison of CORAL on DressCode ( Morelli et al., 2022 ) .

Figure J.13 : Additional Qualitative Comparison of CORAL on PPR10K ( Liang et al., 2021 ) .

Figure J.14 : Additional Qualitative Results of CORAL on PPR10K.

Figure J.15 : Qualitative Comparisons on VITON-HD ( Choi et al., 2021 ) Before and After CORAL Loss.

Figure J.16 : Qualitative Comparisons on DressCode ( Morelli et al., 2022 ) Before and After CORAL Loss.

Figure K.18 : Details on VLM Evaluation Prompt.

Virtual Try-On

Virtual try-on generates images or videos showing people wearing selected garments while preserving appearance, fit, pose, and identity.

See all 55 papers

OmniVTON++: Training-Free Universal Virtual Try-On with Principal Pose Guidance
Feb 16, 2026Zhaotong Yang, Yong Du, Shengfeng He +5Virtual Try-On
School of Computer Science and Technology, Ocean University of China, Qingdao, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China · Sanya Oceanographic Institution, Ocean University of China, Sanya, China +2
FitVTON: Fit-aware Virtual Try-On via Body-Garment Size Control
Jun 10, 2026Yiqun Ning, Ao Shen, Chenhang He +1Virtual Try-On
Department of Computing, The Hong Kong Polytechnic University, Hong Kong · Nuvatech
SIFT-VTON: Geometric Correspondence Supervision on Cross-Attention for Virtual Try-On
May 2, 2026Kosuke Takemoto, Takafumi KoshinakaVirtual Try-On Diffusion-Based Image Generation
The Graduate School of Data Science, Yokohama City University, Kanagawa, Japan

Feb 16, 2026J/K move · Enter open · S save

OmniVTON++: Training-Free Universal Virtual Try-On with Principal Pose Guidance

Zhaotong Yang, Yong Du, Shengfeng He, Yuhui Li+4

School of Computer Science and Technology, Ocean University of China, Qingdao, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China · Sanya Oceanographic Institution, Ocean University of China, Sanya, China+2

Virtual Try-On

Open paper AskPDF arXiv

Image-based Virtual Try-On (VTON) concerns the synthesis of realistic person imagery through garment re-rendering under human pose and body constraints. In practice, however, existing approaches are typically optimized for specific data conditions, making their deployment reliant on retraining and limiting their generalization as a unified solution. We present OmniVTON++, a training-free VTON framework designed for universal applicability. It addresses the intertwined challenges of garment alignment, human structural coherence, and boundary continuity by coordinating Structured Garment Morphing for correspondence-driven garment adaptation, Principal Pose Guidance for step-wise structural regulation during diffusion sampling, and Continuous Boundary Stitching for boundary-aware refinement, forming a cohesive pipeline without task-specific retraining. Experimental results demonstrate that OmniVTON++ achieves state-of-the-art performance across diverse generalization settings, including cross-dataset and cross-garment-type evaluations, while reliably operating across scenarios and diffusion backbones within a single formulation. In addition to single-garment, single-human cases, the framework supports multi-garment, multi-human, and anime character virtual try-on, expanding the scope of virtual try-on applications. The code is available at https://github.com/Jerome-Young/OmniVTON-PlusPlus.

Abstract

Virtual Try-On

Explore similar work

OmniVTON++: Training-Free Universal Virtual Try-On with Principal Pose Guidance

FitVTON: Fit-aware Virtual Try-On via Body-Garment Size Control

SIFT-VTON: Geometric Correspondence Supervision on Cross-Attention for Virtual Try-On