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

Zhaotong Yang; Yong Du; Shengfeng He; Yuhui Li; Xinzhe Li; Yangyang Xu; Junyu Dong; Jian Yang

cs.CVFeb 16, 2026

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

Authors: Zhaotong Yang, Yong Du, Shengfeng He, Yuhui Li, Xinzhe Li, Yangyang Xu, Junyu Dong, Jian Yang

Organizations: 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 · School of Computing and Information Systems, Singapore Management University, Singapore ·

Abstract

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.

Figures & tables

Fig. 1 : We present OmniVTON++, a training-free framework that unifies virtual try-on from in-shop to in-the-wild. It accommodates varied garment sources and person inputs, maintaining consistent garment geometry, texture, and human structure. Notably, OmniVTON++ unlocks multi-garment, multi-human, and anime virtual try-on beyond traditional real-person single-instance settings, all without additional tuning.

Fig. 2 : Overview of OmniVTON++. OmniVTON++ performs Structured Garment Morphing (SGM) to obtain an adapted garment prior, which is incorporated into garment-infused image inpainting, where Principal Pose Guidance (PPG) enforces pose alignment and Continuous Boundary Stitching (CBS) improves boundary consistency across different model backbones. The crossed-circle marker ( $\oslash$ ) indicates components derived from OmniVTON that are not included in the extended journal version.

Fig. 3 : Pose guidance mechanisms in OmniVTON and OmniVTON++. (a) Spectral Pose Injection (SPI) in OmniVTON incorporates human structure only at initialization by combining random noise with an inverted latent. (b) Principal Pose Guidance (PPG) in OmniVTON++ enforces pose-consistent structure throughout diffusion sampling via step-wise noise selection, guided by a proxy latent. (c) A proxy image is constructed via ordered region-wise composition to preserve pose while excluding garment appearance.

Fig. 4 : Cross-stream correspondence in CBS and CBS-DiT on DiT backbones. Person-stream attention is shown as an illustrative example. (a) CBS exhibits positional ambiguity under a shared positional space. (b) CBS-DiT with Positional Index Realignment (PIR) establishes a continuous positional space, yielding aligned correspondence.

person image

I_{p}

; garment image

I_{c}

;

text prompt

c_{txt}

; person mask

M_{p}

garment masks

M_{o}

M_{s}

, body mask

M_{d}

;

diffusion steps

T

; backbone type (U-Net or DiT).

Algorithm 1 OmniVTON++

Virtual Try-on Method
Method	Venue & Year	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$
GP-VTON [ 22 ]	CVPR 2023	51.566	49.196	0.810	0.249
CAT-DM [ 10 ]	CVPR 2024	28.869	26.339	0.775	0.229
D 4 -VTON [ 12 ]	ECCV 2024	25.299	23.914	0.790	0.250
IDM-VTON [ 18 ]	ECCV 2024	23.035	20.460	0.812	0.147
OOTDiffusion [ 4 ]	AAAI 2025	26.312	21.560	0.757	0.208

TABLE I : Quantitative comparisons on VITON-HD under the Shop-to-Model setting. Subscripts

u

and

p

denote the unpaired and paired settings, respectively. The best , second-best , and third-best results are highlighted in each cell.

Method	Venue & Year	DressCode-Upper				DressCode-Lower				DressCode-Dresses
Method	Venue & Year	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$
Virtual Try-On Method
GP-VTON [ 22 ]	CVPR 2023	50.628	47.234	0.879	0.183	102.018	99.136	0.886	0.195	65.028	63.381	0.770	0.274
CAT-DM [ 10 ]	CVPR 2024	14.772	12.362	0.904	0.077	21.990	18.366	0.870	0.116	34.610	32.558	0.800	0.181
D 4 -VTON [ 12 ]	ECCV 2024	20.726	18.243	0.765	0.276	34.088	31.067	0.746	0.284	42.230	41.418	0.764	0.262
IDM-VTON [ 18 ]	ECCV 2024	14.385	12.164	0.895	0.083	21.554	17.777	0.855	0.124	19.745	17.490	0.775	0.207

TABLE II : Quantitative comparisons on DressCode under the Shop-to-Model setting. Subscripts

u

and

p

denote the unpaired and paired settings, respectively. The best , second-best , and third-best results are highlighted in each cell.

	Shop-to-Street	Model-to-Model	Model-to-Street	Street-to-Street
	FID $\downarrow$	FID $\downarrow$	FID $\downarrow$	FID $\downarrow$
Virtual Try-On Method
CAT-DM [ 10 ]	37.484	-	-	-
D 4 -VTON [ 12 ]	35.003	-	-	-
IDM-VTON [ 18 ]	42.282	-	-	-
OOTDiffusion [ 4 ]	42.318	-	-	-

TABLE III : Quantitative comparisons on the StreetTryOn benchmark. The best , second-best , and third-best results are highlighted in each cell.

Fig. 5 : Qualitative results under cross-dataset and cross-garment-type evaluation. Upper-garment try-on results on VITON-HD are shown on the top, while lower-garment and dress results on DressCode are shown on the bottom.

Fig. 6 : Qualitative results on the StreetTryOn benchmark across different scenarios. From top to bottom: Shop-to-Street, Model-to-Model, Model-to-Street, and Street-to-Street.

Method	Backbone	SGM	PPG	CBS	CBS-DiT	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$
Base	SD-2.0					18.445	16.878	0.773	0.222
(A)	SD-2.0	✓				12.946	10.949	0.803	0.173
(B)	SD-2.0		✓			12.319	10.604	0.827	0.167
(C)	SD-2.0	✓		✓		9.331	7.484	0.821	0.154
Ours	SD-2.0	✓	✓	✓		9.189	6.990	0.843	0.130
Base*	FLUX					13.900	11.564	0.827	0.167

TABLE IV : Macro ablation of Structured Garment Morphing (SGM), Continuous Boundary Stitching (CBS), and Spectral Pose Injection (SPI) on VITON-HD under the SD-2.0 and FLUX backbones. Best results are marked in bold per block.

Variant	Backbone	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$
Outpainting-based [ 9 ]	SD-2.0	9.470	7.429	0.842	0.135
DensePose Flow-based	SD-2.0	12.043	10.540	0.840	0.145
SGM (OmniVTON++)	SD-2.0	9.189	6.990	0.843	0.130

TABLE V : Ablation study on alternative designs for SGM. Subscripts

u

and

p

denote the unpaired and paired settings, respectively. The best results are marked in bold .

Fig. 7 : Qualitative ablation results of component variants under SD-2.0 (left) and FLUX (right). Compared with individual or partial configurations, the complete model yields more accurate garment geometry and more coherent body structure.

Fig. 8 : Qualitative comparisons of pseudo-person image generation and downstream try-on results using attention-modulated outpainting (OmniVTON) and virtual dressing (OmniVTON++).

Variant	Backbone	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$
ControlNet [ 58 ]	SD-2.0	10.910	8.778	0.818	0.159
SPI [ 9 ]	SD-2.0	9.365	6.964	0.832	0.141
Full-Latent	SD-2.0	9.462	7.244	0.847	0.128
Low-Frequency Latent	SD-2.0	9.514	7.431	0.835	0.139
PPG (OmniVTON++)	SD-2.0	9.189	6.990	0.843	0.130

TABLE VI : Ablation study on pose guidance designs. Subscripts

u

and

p

denote the unpaired and paired settings, respectively. The best and second-best results are marked in bold and underlined .

Fig. 9 : Visualization of principal and low-frequency latent components. The principal components reveal a more compact encoding of pose-related structure.

Variant	Backbone	FID u $\downarrow$	FID p $\downarrow$	SSIM p $\uparrow$	LPIPS p $\downarrow$
Continuous Boundary Stitching (CBS)
Attention Reweighting [ 60 ]	SD-2.0	13.143	11.048	0.823	0.156
Delta Denoising Score [ 61 ]	SD-2.0	12.125	10.149	0.829	0.154
Asymmetric Interaction	SD-2.0	9.324	7.188	0.841	0.131
CBS (OmniVTON++)	SD-2.0	9.189	6.990	0.843	0.130
Continuous Boundary Stitching for DiT (CBS-DiT)

TABLE VII : Ablation study on alternative designs for CBS and CBS-DiT. Subscripts

u

and

p

denote the unpaired and paired settings, respectively. Best results are marked in bold per block.

Fig. 10 : Qualitative results of multi-garment virtual try-on with OmniVTON++.

Fig. 11 : Multi-human virtual try-on results. Top row shows Model-to-Model, while bottom row depicts Shop-to-Street.

Fig. 13 : Failure cases of OmniVTON++. (a) Imprecise human analysis signals. (b) Incorrect accessory masking.

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

OmniDiT: Extending Diffusion Transformer to Omni-VTON Framework
Mar 20, 2026Weixuan Zeng, Pengcheng Wei, Huaiqing Wang +8Virtual Try-On Diffusion Transformer
The Chinese University of Hong Kong, ShenZhen · Beihang University · KuaiShou
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
TAMF-VTON: Token-Wise Sparse Adaptation for Mask-Free Virtual Try-On
Jul 16, 2026Jie Wang, Qian He, Gaofeng He +2Virtual Try-On Compositional Image Generation
State Key Lab of CAD & CG, Zhejiang University and Style3D Research, China · Style3D Research, China · State Key Lab of CAD & CG, Zhejiang University, China

Mar 20, 2026J/K move · Enter open · S save

OmniDiT: Extending Diffusion Transformer to Omni-VTON Framework

Weixuan Zeng, Pengcheng Wei, Huaiqing Wang, Boheng Zhang+7

The Chinese University of Hong Kong, ShenZhen · Beihang University · KuaiShou

Virtual Try-On Diffusion Transformer Efficient Transformer Inference

Open paper AskPDF arXiv

Despite the rapid advancement of Virtual Try-On (VTON) and Try-Off (VTOFF) technologies, existing VTON methods face challenges with fine-grained detail preservation, generalization to complex scenes, complicated pipeline, and efficient inference. To tackle these problems, we propose OmniDiT, an omni Virtual Try-On framework based on the Diffusion Transformer, which combines try-on and try-off tasks into one unified model. Specifically, we first establish a self-evolving data curation pipeline to continuously produce data, and construct a large VTON dataset Omni-TryOn, which contains over 380k diverse and high-quality garment-model-tryon image pairs and detailed text prompts. Then, we employ the token concatenation and design an adaptive position encoding to effectively incorporate multiple reference conditions. To relieve the bottleneck of long sequence computation, we are the first to introduce Shifted Window Attention into the diffusion model, thus achieving a linear complexity. To remedy the performance degradation caused by local window attention, we utilize multiple timestep prediction and an alignment loss to improve generation fidelity. Experiments reveal that, under various complex scenes, our method achieves the best performance in both the model-free VTON and VTOFF tasks and a performance comparable to current SOTA methods in the model-based VTON task.

Abstract

Virtual Try-On

Explore similar work

OmniDiT: Extending Diffusion Transformer to Omni-VTON Framework

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

TAMF-VTON: Token-Wise Sparse Adaptation for Mask-Free Virtual Try-On