cs.CVOct 29, 2026

FlexTry-On: A Time-Efficient Framework for Multi-Scenario Virtual Try-on

Authors: Phuong Dam, Quyen Nguyen Duc, Son Hoang Minh, Daeyoung Kim

Organizations: Korea Advanced Institute of Science and Technology (KAIST)

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Virtual try-on has achieved remarkable progress in recent years, yet most existing methods are designed for a specific scenario, such as cloth-to-person or person-to-person try-on, and often require separate conditioning strategies and processing pipelines for different settings. We present FlexTry-On, a unified and time-efficient diffusion-based framework for diverse virtual try-on scenarios, including cloth-to-person, person-to-person, cross-domain, and in-the-wild settings. FlexTry-On adopts a common conditioning formulation, where a text-aware condition extraction strategy converts heterogeneous garment sources into a unified representation, and a dual-conditioning diffusion transformer jointly models garment appearance and person identity. Extensive experiments demonstrate competitive performance on both cloth-to-person and person-to-person benchmarks while maintaining fast inference. Moreover, the same framework can be directly applied to unseen cross-domain scenarios without requiring separate architectures or scenario-specific processing pipelines. These results highlight the potential of unified and efficient virtual try-on frameworks as a practical step toward more flexible and widely applicable virtual try-on systems.

Figures & tables

Figure 2: Comparison of FlexTry-On with other diffusion-based methods using the VITON-HD dataset (C2P). Lower values indicate better performance for both FID and inference speed. FlexTry-On supports all evaluated virtual try-on scenarios within a unified framework, as summarized in Tab. 1 .
Figure 3: Overview of the FlexTry-On framework for IVTON. Multiple reference inputs, including pose keypoints JpJ_{p} and DensePose maps IdpI_{dp} , are encoded into latent representations and concatenated as inputs to the diffusion process. The TASM module receives a latent image with text conditions embedded via the CLIP-Text Encoder [ 39 ] from SD 2.0. The TASM outputs both the latent representations of the target RGB regions and their corresponding masks. In TASM, we illustrate the latent representation of the RGB input cloth PicP_{ic} and its associated mask MicM_{ic} . The garment-agnostic mask MgaM_{ga} is obtained by merging masks under different text conditions according to Eq. 2
Figure 4: Overview of the Dual-DiT Tryon Module (DDTM). The DDTM follows a DiT-based architecture [ 35 ] with cross-attention to integrate person and garment information. During training, different conditioning strategies are adopted for C2P and P2P settings. For clarity, patchification and positional embeddings are omitted.
 Inf. TimeVITON-HDDress Code - UpperSupport IVTON Tasks
Method(s) ↓\downarrowFID (pair/unp.) ↓\downarrowSSIM ↑\uparrowLPIPS ↓\downarrowFID (pair/unp.) ↓\downarrowSSIM ↑\uparrowLPIPS ↓\downarrowC2PP2PCross C2PCross P2P
MDNM †\dagger [IJCAI 2025] [ 9 ]--/9.380.8720.086---✓✗✗✗
APM †\dagger [ICCV 2025] [ 10 ]--/9.380.9010.079---✓✗✗✗
M&MVTO [CVPR 2024] [ 48 ]-------✓✓✗✗
OmniVTON †\dagger [ICCV 2025] [ 44 ]1407.76/9.620.8580.098---✓✓✗✗
LaDI-VTON [ACM 2023] [ 32 ]2.19.63/12.310.8580.09810.45/14.050.9270.062✓✗✗✗
Table 1: Quantitative comparison on VITON-HD and DressCode Upper in C2P. Inf. Time denotes method-specific inference time; shared pose/DensePose preprocessing is excluded, with exceptions detailed in Appendix E . FlexTry-On is trained only on VITON-HD and DressCode, whereas FlexTry-On* uses the full uncurated C2P–P2P corpus. Best/second-best results are bold / underlined . Cross C2P/P2P denote cross-domain variants. DressCode Lower/Dresses results are in Appendix B . †\dagger : paper-reported.
Figure 5: An example of mislabeled pairs in the IMAGDress dataset [ 40 ] . The black arrow indicates the correct pairing, whereas the dataset incorrectly associates each person’s image with all garment images, as shown by red dashed arrows.
Figure 6: Cross-domain IVTON results on DressCode [ 33 ] . Existing methods [ 4 , 3 , 19 , 16 , 8 , 22 , 23 , 32 ] typically rely on category-specific human parsing [ 28 ] or human-centric segmentation models such as Sapiens [ 21 ] (e.g., Voost [ 26 ] ), which become ambiguous when defining editable regions in cross-domain scenarios. In contrast, our framework employs TASM for text-conditioned segmentation, enabling task-aware mask generation across different garment categories and supporting cross-domain IVTON. The TASM module operates in latent space conditioned on text; masks are interpolated for visualization purposes.
Figure 7: Qualitative comparison with state-of-the-art methods on VITON-HD [ 2 ] . All methods are evaluated on the same test inputs using their publicly released checkpoints when available. Conventional human parsing [ 28 ] may introduce inaccurate garment boundaries and incomplete editable regions, which can affect downstream virtual try-on performance. In contrast, TASM generates task-aware masks specifically designed for virtual try-on conditioning, resulting in improved garment preservation and reduced visual artifacts. Note that the TASM module operates in latent space conditioned on text; masks are interpolated for visualization purposes.
Figure 8: Qualitative results in C2P and P2P IVTON setting. The person and clothes images are collected from the Internet. FlexTry-On* is our proposed method trained on merged uncurated datasets. ITA-MDT* denotes the fine-tuned ITA-MDT with segmented clothing, trained on our dataset (the full, uncurated dataset).
 Seg ParamTime / try-onVITON-HDDressCode - Upper
Method(M Param) ↓\downarrowSeg. (s) ↓\downarrowDenoising (s) ↓\downarrowFID (pair/unp.) ↓\downarrowSSIM ↑\uparrowLPIPS ↓\downarrowFID (pair/unp.) ↓\downarrowSSIM ↑\uparrowLPIPS ↓\downarrow
DDTM (Baseline)650.11.16.21/9.640.8570.0918.89/12.840.9110.061
Baseline + TASM1300.11.15.83 / 9.020.8610.0857.12 / 11.110.9250.042
Baseline + TASM + Lstyle\mathcal{L}_{style}1300.11.15.51 / 8.970.8610.0867.15 / 11.070.9270.042
Baseline + G-SAM13619.01.17.15/10.270.8410.107---
DDTM* + TASM1300.11.16.80/9.420.8570.0938.22/11.920.9230.048
Table 2: Ablation study quantitative results on VITON-HD [ 2 ] and DressCode [ 33 ] Upper in C2P. DDTM denotes the DDTM trained only on VITON-HD and DressCode in C2P, whereas DDTM* denotes training on the full uncurated combined corpus. Best/second-best results are bold / underlined . Times are measured for one batch-size-1 try-on. For that image, each segmenter evaluates all 10 segmentation-class conditions listed in Appendix D . TASM batches them in 0.1 s, whereas G-SAM runs GroundingDINO followed by SAM sequentially for each condition (1.9 s per class; 19.0 s total). Denoising takes 1.1 s per try-on.
Figure 9: Qualitative StreetTryOn results [ 6 ] . * marks our unsuccessful reproduction.
 Shop2StreetModel2ModelModel2StreetStreet2Street
MethodFID↓FID↓FID↓FID↓
CAT-DM †\dagger [CVPR 2024] [ 45 ]37.48---
IDM-VTON †\dagger [ECCV 2024] [ 3 ]42.28---
OmniVTON †\dagger [ICCV 2025] [ 44 ]33.928.9833.4523.47
FlexTry-On (Ours)32.5110.0633.0732.6
Table 3: Results on Street Try-On. †\dagger : OmniVTON [ 44 ] reported the result
Figure 10: Qualitative ablation on VITON-HD [ 2 ] under curated and full noisy-data training. DDTM and DDTM* denote the two settings, respectively; “Use G-SAM” replaces TASM with Grounded-SAM under matched-text conditions.
Figure 11: Qualitative result on VITON-HD, P2P setting. Model*: retrained model.
 Inf. TimeVITON-HD P2PVITON-HD C2P
Method(s) ↓\downarrowFID (pair/unp.) ↓\downarrowSSIM ↑\uparrowLPIPS ↓\downarrowFID (pair/unp.) ↓\downarrowSSIM ↑\uparrowLPIPS ↓\downarrow
Stable VITON* [CVPR 2024] [ 22 ]6.87.01 / 10.350.8760.0986.93 / 9.530.8560.091
CAT-DM* [CVPR 2024] [ 45 ]3.510.71/12.440.8320.1538.72/11.320.8430.124
ITA-MDT* [CVPR 2025] [ 19 ]3.28.02/11.910.8810.0787.99/11.340.8760.094
FlexTry-On* (Ours)1.24.01 / 10.060.9050.0446.34 / 9.220.8590.093
Table 4: P2P and C2P results on VITON-HD. ∗ denotes retraining on our full C2P–P2P corpus. Best and second-best results are bold and underlined .
 VITON-HDDress Code - Upper
MethodFID (pair/unp.) ↓SSIM ↑LPIPS ↓FID (pair/unp.)↓SSIM ↑LPIPS ↓
DDTM + TASM-MaskOnly + Lstyle\mathcal{L}_{style}5.98/9.080.8600.0867.20/11.080.9240.043
DDTM + TASM + Lstyle\mathcal{L}_{style}5.51 / 8.970.8610.0867.15 / 11.070.9270.042
Table A : Quantitative results of diffusion-based virtual try-on methods on VITON-HD and DressCode upper-body test sets. We report both paired and unpaired FID. TASM-MaskOnly denotes the TASM with output that is the mask only. DDTM refers to the training of the Dual-DiT Tryon Module with only VITON-HD [ 2 ] and Dresscode [ 33 ] in the C2P tasks. The better performance for each metric is bold .
Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A : Overview of the Text-aware segmentation Module for FlexTry-On Framework and its variant. The TASM module receives a latent image with text conditions embedded via the CLIP-Text Encoder from Stable Diffusion 2.0 [ 39 ] . The module outputs both the latent representations of the target RGB regions and their corresponding masks. In TASM, we illustrate the latent representation of the RGB input cloth PicP_{ic} and its associated mask MicM_{ic} . While in TASM-MaskOnly, the output is only the mask MicM_{ic} .
Figure B : Ablation study results on VITON-HD [ 2 ] to evaluate the effectiveness of the TASM module architecture design in the clean datasets. The DDTM denotes the Dual-DiT Tryon Module, which is trained only on the clean dataset (VITON-HD [ 2 ] and DressCode [ 33 ] ).
DatasetResolutionsC2PP2P# of images
VITON-HD [ 2 ]1024 x 768✓✗11 647
DressCode [ 33 ]1024 x 768✓✗53 792
DeepFashion [ 31 ]Varies✗✓52 712
IMAGDress [ 40 ]Varies✓✓324 857
LRVS-Fashion [ 27 ]Varies✓✓500 000
Appendix
Table C : IVTON Data Statistic
Figure C : Representative Street2Street failure cases on Street Try-On [ 6 ] . FlexTry-On may preserve the garment’s appearance while failing to fully reproduce its fine-grained structure.
Figure D : Additional qualitative results on heterogeneous subjects with varied appearances, body shapes, poses, framing, and backgrounds. All outputs are generated using the same FlexTry-On* checkpoint without additional adaptation.
Figure E : Qualitative results on DressCode [ 33 ] Lower. Our model (Ours) here is trained only with VITON-HD [ 2 ] and DressCode [ 33 ] .
Figure F : Qualitative results on DressCode [ 33 ] Dresses. Our model (Ours) here is trained only with VITON-HD [ 2 ] and DressCode [ 33 ] .
Figure G : Qualitative results on Cross Domain (Upper-Lower to Dress) with C2P and P2P Setting. The dresses are taken from DressCode [ 33 ] , and the person’s images are taken from the Internet. FlexTry-On* refers to our training on the combined uncurated datasets. ITA-MDT* [ 19 ] denotes the fine-tuned ITA-MDT with segmented clothing, trained on our dataset.
Figure H : Qualitative results on Cross Domain (Dress to Upper) with C2P Setting. The Upper Garment and person’s image are taken from DressCode [ 33 ] . ITA-MDT* [ 19 ] denotes the fine-tuned ITA-MDT with segmented clothing, trained on our dataset. FlexTry-On* refers to training on the combined uncurated datasets. Existing methods [ 4 , 3 , 19 , 16 , 8 , 22 , 23 , 32 ] based on conventional human parsing [ 28 ] are limited in supporting flexible masking for cross-domain try-on. In this experiment, we either masked all clothing or used full-body clothing masks for ITA-MDT and CatVTON.
Figure I : Qualitative results on Cross Domain (Dress to Upper) with P2P Setting. The Upper Garment and person’s image are taken from DressCode [ 33 ] . ITA-MDT* [ 19 ] denotes the fine-tuned ITA-MDT with segmented clothing, trained on our dataset. FlexTry-On* refers to training on the combined uncurated datasets. Existing methods [ 4 , 3 , 19 , 16 , 8 , 22 , 23 , 32 ] typically rely on category-specific masks derived from conventional human parsing pipelines [ 28 ] . In cross-domain scenarios, however, the editable regions cannot be uniquely determined from predefined garment categories. For ITA-MDT and CatVTON, we therefore evaluate two representative masking strategies: masking all clothing regions and using full-body clothing masks. In contrast, TASM generates task-aware masks conditioned on the target garment description.
Figure J : Qualitative results on Cross Domain (Dress to Lower) with C2P Setting. The Lower Garment and person’s image are taken from DressCode [ 33 ] . ITA-MDT* [ 19 ] denotes the fine-tuned ITA-MDT with segmented clothing, trained on our dataset. FlexTry-On* refers to training on the combined uncurated datasets. Existing methods [ 4 , 3 , 19 , 16 , 8 , 22 , 23 , 32 ] typically rely on category-specific masks derived from conventional human parsing pipelines [ 28 ] . In cross-domain scenarios, however, the editable regions cannot be uniquely determined from predefined garment categories. For ITA-MDT and CatVTON, we therefore evaluate two representative masking strategies: masking all clothing regions and using full-body clothing masks. In contrast, TASM generates task-aware masks conditioned on the target garment description.
Figure K : Qualitative results on Cross Domain (Dress to Lower) with P2P Setting. The Lower Garment and person’s image are taken from DressCode [ 33 ] . ITA-MDT* [ 19 ] denotes the fine-tuned ITA-MDT with segmented clothing, trained on our dataset. FlexTry-On* refers to training on the combined uncurated datasets. Existing methods [ 4 , 3 , 19 , 16 , 8 , 22 , 23 , 32 ] typically rely on category-specific masks derived from conventional human parsing pipelines [ 28 ] . In cross-domain scenarios, however, the editable regions cannot be uniquely determined from predefined garment categories. For ITA-MDT and CatVTON, we therefore evaluate two representative masking strategies: masking all clothing regions and using full-body clothing masks. In contrast, TASM generates task-aware masks conditioned on the target garment description.
More in this topic

Virtual Try-On

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

See all 56 papers

Explore similar work

CardsList
Feb 16, 2026

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

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.

Zhaotong Yang, Yong Du, Shengfeng He +5

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
AskPDF
Jun 7, 2026

OmniTryOn: Video Try-On Anything at Once!

Although video virtual try-on (VVT) has achieved significant progress, existing methods still exhibit two fundamental limitations: first, they are restricted to single-garment transfer, rendering simultaneous multi-object try-on highly impractical; second, their heavy reliance on explicit external priors (e.g., garment masks) inevitably destroys crucial physical dynamics and degrades visual quality. To bridge this gap, this paper proposes the novel Try-On Anything task, which aims to simultaneously transfer diverse wearable objects onto a person in a video in a single inference pass. To support and standardize this paradigm, we introduce TryAny-Bench, a comprehensive benchmark encompassing a paired video dataset alongside a tailored evaluation protocol. Furthermore, we present OmniTryOn, an external-prior-free generative framework designed to tackle this task. Specifically, OmniTryOn employs a First Frame Wearable Cache strategy, which directly provides diverse wearable objects for the generation process through the initial video frame. To maintain consistency, we propose the Spatiotemporally Consistent RoPE (STC-RoPE), which inherently establishes robust spatiotemporal anchors to strictly preserve complex human motions and background dynamics. Optimized by the proposed Gradual Try-On (GTO) training strategy, our model progressively masters robust multi-object synthesis. Extensive experiments on TryAny-Bench demonstrate that OmniTryOn significantly outperforms existing specialized video virtual try-on models and general video editing baselines, establishing a powerful new standard for the Try-On Anything task. Our dataset, code, and models are available at https://github.com/xcltql666/OminTryOn.

Changliang Xia, Chengyou Jia, Minnan Luo +3

Xi’an Jiaotong University Xi’an, China

Virtual Try-OnTemporal Consistency in Video Generation
AskPDF
Jul 23, 2026

Oxygen-TryOn: Fashion-Native Foundation Model for Any-item Virtual Try-On

We present Oxygen-TryOn, a unified foundation model for any-item virtual try-on. Rather than repurposing a general-purpose image editor, Oxygen-TryOn is fashion-native, built for try-on through a dedicated data engine and try-on-specific training. Given one or more reference items (clean product shots or in-the-wild worn-on photos) and a single target subject image, it synthesizes a photorealistic image of the subject wearing the items across virtually any fashion category. Prior systems handle a single garment category in a studio setting, and recent multi-reference methods remain garment-centric; in contrast, Oxygen-TryOn supports diverse items and scenarios, including full- and half-body views, a variable number of references, and free multi-item composition, while faithfully preserving both subject identity and item appearance. Instead of mask-based inpainting, we reformulate try-on as a multi-reference, understanding-driven generation task. We build a data engine that collects, manufactures, annotates, and filters high-quality try-on data at scale, and design a three-stage recipe of continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL). The RL stage uses a hybrid reward combining an in-house try-on reward model with a proprietary, rubric-guided general-purpose model, jointly supervising fine-grained consistency and instruction-level quality. It also follows general editing instructions (e.g., pose changes) in the same pass. Across public benchmarks and our in-house Oxygen-TryOn Bench, it achieves state-of-the-art consistency and realism on single-item try-on and leads on multi-item try-on, matching or surpassing both leading proprietary systems (Nano Banana Pro, GPT-Image-2, Seedream5 Lite) and open-source models (FLUX.2).

Yong Liu, Xiaolong Fu, Zihang Xu +10

Oxygen AIGC Group & Joy Future Academy, JD

Virtual Try-OnConditional Image GenerationImage Generation
AskPDF