FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On

Johanna Karras; Yuanhao Wang; Yingwei Li; Ira Kemelmacher-Shlizerman

cs.CVApr 9, 2026

FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On

Authors: Johanna Karras, Yuanhao Wang, Yingwei Li, Ira Kemelmacher-Shlizerman

Organizations: Paul G. Allen School for Computer Science and Engineering, University of Washington, Seattle, WA, USA · Google Research, Seattle, CA, USA · Google Research, Mountain View, CA, USA

Abstract

Given a person and a garment image, virtual try-on (VTO) aims to synthesize a realistic image of the person wearing the garment, while preserving their original pose and identity. Although recent VTO methods excel at visualizing garment appearance, they largely overlook a crucial aspect of the try-on experience: the accuracy of garment fit -- for example, depicting how an extra-large shirt looks on an extra-small person. A key obstacle is the absence of datasets that provide precise garment and body size information, particularly for "ill-fit" cases, where garments are significantly too large or too small. Consequently, current VTO methods default to generating well-fitted results regardless of the garment or person size. In this paper, we take the first steps towards solving this open problem. We introduce FIT (Fit-Inclusive Try-on), a large-scale VTO dataset comprising over 1.13M try-on image triplets accompanied by precise body and garment measurements. We overcome the challenges of data collection via a scalable synthetic strategy: (1) We programmatically generate 3D garments using GarmentCode and drape them via physics simulation to capture realistic garment fit. (2) We employ a novel re-texturing framework to transform synthetic renderings into photorealistic images while strictly preserving geometry. (3) We introduce person identity preservation into our re-texturing model to generate paired person images (same person, different garments) for supervised training. Finally, we leverage our FIT dataset to train a baseline fit-aware virtual try-on model. Our data and results set the new state-of-the-art for fit-aware virtual try-on, as well as offer a robust benchmark for future research. We will make all data and code publicly available on our project page: https://johannakarras.github.io/FIT.

Figures & tables

Figure 1. The FIT Dataset. We present FIT, a dataset and benchmark designed for fit-aware virtual try-on, featuring diverse garment fits (e.g., tight, loose) and precise size annotations. Left: Sample dataset triplets showing the conditioning garment image ( top ), the conditioning person image ( middle ), and the target try-on image ( bottom ). Right: Visualization of the corresponding person and garment measurement annotations. Backgrounds are removed for clarity. Teaser.

Figure 2. FIT dataset generation. (a) Overall pipeline: For each sample, we first simulate garment draping in 3D via GarmentCode, rendering a synthetic try-on image $I_{s}$ (see (b)). Then, we generate a text prompt $p$ (via VLM) describing the person and garment appearance, as well as a composite normal map $I_{n}$ based on $I_{s}$ . We use $p$ and $I_{n}$ to condition our re-texturing model $f_{\text{texture}}$ to generate the photorealistic try-on image $I_{\text{try-on}}$ . Finally, our model $f_{\text{paired}}$ generates a paired person image $I_{p}$ (see (c)) and a VLM synthesizes the corresponding layflat garment $I_{g}$ . (b) GarmentCode simulation: Given a garment design template, we compute a sewing pattern with measurements $m_{g}$ for a specific body size $A$ . Then, we cross-drape the pattern onto a different target body of size $B$ with person measurements $m_{p}$ , using box-mesh realignment to prevent simulation failures. (c) To generate a paired person image (same person and pose, different garment), we start with a paired rendered image $I_{s}^{\prime}$ containing a different garment than in $I_{\text{try-on}}$ draped onto the same body. Next, we derive an identity map $I_{\text{id}}$ by masking out the combined source and paired garment regions in $I_{\text{try-on}}$ . Conditioned on $I_{\text{id}}$ , the paired normal map $I_{n}^{\prime}$ , and a paired prompt $p^{\prime}$ , $f_{\text{paired}}$ generates $I_{p}$ . FIT data generation pipeline.

Dataset	Realism	Ill-Fit	Measurements	Triplet	Scale
SV-VTO	✓	✓	✓	✓	1,524
SIZER	✓	✓	✓	✗	2,000
DeepFashion3D	✗	✗	✗	✗	2,078
ViTON-HD	✓	✗	✗	✗	11,647
Size4Men	✓	✓	✗	✗	13,000
LAION-Garment	✓	✗	✗	✓	60K

Table 1. Comparison of related datasets. We compare FIT to several related datasets. For scale, we report the number of training images.

Figure 3. Fit-VTO architecture. Our architecture is a flow-based diffusion model based on Flux.1-dev ( Black Forest Labs, 2024 ) and finetuned with LoRA ( Hu et al., 2023 ) . FiT-VTO generates a try-on image $I_{\text{try-on}}$ given a layflat garment image $I_{g}$ , paired person image $I_{p}$ , and person-garment measurements $m=[m_{p},m_{g}]$ . First, image inputs $I_{g}$ and $I_{p}$ are encoded into latents separately through a pre-trained VAE encoder. We replace the text embeddings in Flux.1-dev with custom measurement embeddings $m_{\text{embed}}$ computed from $m$ . Person latents are channel-concatenated with the noisy target latents, while layflat latents and $m_{\text{embed}}$ are sequence-wise concatenated with $z_{t}$ . After processing through the diffusion transformer, clean latents are decoded by the VAE decoder. Fit-VTO architecture.

Figure 4. Paired Image Generation Comparison. VLM methods struggle with pose and shape preservation, while VTO and inpainting baselines introduce artifacts. Our approach yields highly consistent paired data. Paired Image Generation Comparison.

	VITON-HD				FIT Dataset
	SSIM $\uparrow$	FID $\downarrow$	LPIPS $\downarrow$	KID $\downarrow$	SSIM $\uparrow$	FID $\downarrow$	LPIPS $\downarrow$	KID $\downarrow$	IOU $\uparrow$
Any2AnyTryon ( Guo et al., 2025 )	0.758	14.186	0.152	2.413	0.819	25.059	0.209	3.939	0.783
Nano Banana Pro ( Google, 2025a )	0.552	11.344	0.501	0.624	0.785	19.926	0.166	1.676	0.792
COTTON ( Chen et al., 2023 )	0.615	39.117	0.349	11.397	0.759	29.716	0.207	6.269	0.739
IDM-VTON ( Choi et al., 2024 )	0.849	9.115	0.077	0.471	0.739	31.229	0.246	6.819	0.789
Ours ${}_{\text{no FIT}}$	0.817	11.499	0.103	0.639	0.852	16.427	0.095	0.849	0.844

Table 2. Quantitative comparisons. We compare Fit-VTO to related methods and ablated versions of our method. Ours

{}_{\textbf{ft\_vitonhd}}

refers to our method finetuned with VITON-HD training data. Bolded and underlined values indicate the best and second-best scores per column, respectively.

Figure 5. Qualitative results. We show examples of Fit-VTO results on the synthetic FIT test dataset. Fit-VTO respects person and garment inputs, while also synthesizing realistic garment fit based on person and garment measurements (zoom in for details). For brevity, we approximate the full measurements with a size label (XS-3XL). See the appendix for our size categorization chart. Qualitative Fit-VTO results.

Figure 6. Qualitative comparisons. We compare Fit-VTO to related and ablated methods using synthetic FIT test data (top two rows) and real-world VITON-HD data (bottom two rows). Overall, our method best depicts the most accurate garment appearance and fit. Qualitative Fit-VTO comparisons.

Figure 7. Independent size control. Fit-VTO realistically visualizes garment fit across various sizes on a fixed person size. Independent size control.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

	Men’s	Women’s
Measurement	(min, mean, max, std)	(min, mean, max, std)
Bust	(87, 101, 141, 10)	(83, 100, 136, 13)
Height	(155, 174, 194, 8.0)	(151, 170, 196, 9.0)
Hips	(88, 101, 125, 6.0)	(89, 104, 127, 10)
Waist	(70, 86, 141, 13)	(61, 85, 130, 17)

Appendix

Table 3. Body size statistics. We report the min, mean, max, and standard deviation of our garment measurements in cm.

	Men’s	Women’s
Measurement	(min, mean, max, std)	(min, mean, max, std)
Width	(77, 112, 169, 16)	(75, 110, 169, 16)
Length	(29, 53, 76, 8.0)	(29, 51, 76, 7.5)
Sleeve Length	(0.0, 30, 79, 17)	(0, 29, 79, 17)

Appendix

Table 4. Garment size statistics. We report the min, mean, max, and standard deviation of our garment measurements in cm.

	Bust		Waist		Hips
	Men’s	Women’s	Men’s	Women’s	Men’s	Women’s
Size	(min, max)	(min, max)	(min, max)	(min, max)	(min, max)	(min, max)
XS	(86, 91)	(79, 84)	(71, 76)	(58, 64)	(91, 96)	(84, 89)
S	(91, 96)	(86, 89)	(76, 81)	(66, 67)	(96, 101)	(91, 94)
M	(96, 101)	(90, 95)	(81, 86)	(71, 75)	(101, 106)	(97, 102)
L	(101, 106)	(96, 104)	(86, 91)	(86, 91)	(106, 111)	(106, 111)

Appendix

Table 5. Body size categorization statistics. We report the min and max for each body measurements and size label in cm.

Figure 8. During cross-body draping, the initial boxmeshes are often misaligned with the target human models, causing draping failures (left). We explicitly realign the boxmesh to ensure successful simulation (right).

Figure 9. Additional dataset examples. In each example, we show the paired person image (top left), garment image (lower left), and the target try-on image (right).

Figure 10. Qualitative resizing results. Fit-VTO adapts garment fit according to individual garment measurements. We show results of independently shrinking and growing the length, width, and sleeve length with respect to the original value. Please zoom in for details. Qualitative resizing results.

Figure 11. Failure cases. (a) GarmentCode ( Korosteleva and Sorkine-Hornung, 2023 ) simulation does not model varying degrees of tightness well, leading to similar-looking fit for all garment sizes smaller than the body size. (b) As a result of (a), it is difficult to tell the level of tightness in FIT try-on images. (c) Due to the correlations in measurements across sizes, adjustments to individual measurements may also lead to undesired changes in other dimensions. In this example, increasing garment length also increases garment width.

Figure 12. Garment fit distribution. We plot the frequency of each (body size, garment size) pairing in our dataset according to the size classification in Table 5 .

Figure 13. Real-world resizing results. We show Fit-VTO try-on performance on real-world person images using varying garment sizes. Fit-VTO realistically shrinks and grows the garment fit according to uniform adjustments to the garment measurements–length, width, and sleeve length–with respect to their original values (1.0x). The size label to the left of each example corresponds to the person’s body size. Since real-world garment images with precise measurements are difficult to acquire, we use our synthetic garment images and measurements.

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

Jun 10, 2026

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

While diffusion-based virtual try-on has achieved impressive visual realism, most methods treat the task as 2D inpainting, prioritizing texture preservation over physical plausibility. Consequently, they often produce plausible-looking images that fail to reflect authentic garment fit across diverse body shapes. We present FitVTON, a Fit-aware virtual try-on model on different bodies in the wild. FitVTON encodes garment-body size through structured text prompts, and learn from simulated try-on triplets from parameterized garment model. To improve the fitting effects over garment silhouettes, we introduce two auxiliary head to predict the masks for both the garment and the exposed body. We further introduce a texture rectification stage to improve realistic appearance from simulated data. To evaluate the fitting fidelity, we curate a real-world dataset, FittingEffect3K, combining VLM-based scoring protocol. Both subjective and quantitive experiments show that FitVTON demonstrate authentic fitting fidelity, with significant sizing accuracy and shape preservation over state-of-the-art methods while maintaining competitive image quality. Project Page: https://zenoning.github.io/FitVTON/.

Yiqun Ning, Ao Shen, Chenhang He +1

Department of Computing, The Hong Kong Polytechnic University, Hong Kong · Nuvatech

Virtual Try-On

Ask PDF

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

Ask PDF

Mar 14, 2026

Garments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories

Virtual try-on (VTON) has advanced single-garment visualization, yet real-world fashion centers on full outfits with multiple garments, accessories, fine-grained categories, layering, and diverse styling, remaining beyond current VTON systems. Existing datasets are category-limited and lack outfit diversity. We introduce Garments2Look, the first large-scale multimodal dataset for outfit-level VTON, comprising 80K many-garments-to-one-look pairs across 40 major categories and 300+ fine-grained subcategories. Each pair includes an outfit with 3-12 reference garment images (Average 4.48), a model image wearing the outfit, and detailed item and try-on textual annotations. To balance authenticity and diversity, we propose a synthesis pipeline. It involves heuristically constructing outfit lists before generating try-on results, with the entire process subjected to strict automated filtering and human validation to ensure data quality. To probe task difficulty, we adapt SOTA VTON methods and general-purpose image editing models to establish baselines. Results show current methods struggle to try on complete outfits seamlessly and to infer correct layering and styling, leading to misalignment and artifacts.

Junyao Hu, Zhongwei Cheng, Waikeung Wong +1

The Hong Kong Polytechnic University · Huhu AI Inc.

Virtual Try-On

Ask PDF

Abstract

Virtual Try-On

Explore similar work

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

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

Garments2Look: A Multi-Reference Dataset for High-Fidelity Outfit-Level Virtual Try-On with Clothing and Accessories