MV-Fashion: Towards Enabling Virtual Try-On and Size Estimation with Multi-View Paired Data

Hunor Laczkó; Libang Jia; Loc-Phat Truong; Diego Hernández; Sergio Escalera; Jordi Gonzalez; Meysam Madadi

Table 1: We compare the proposed MV-Fashion dataset with existing datasets, which we divide into two categories: (a) multi-view human, and (b) fashion-centric datasets. For the multi-view human comparison, we evaluate attributes such as Size (inclusion of clothing measurements), Styles (style variations of the same outfit), Layered (multiple clothing layers in the same outfit), Subjects (number of participants), and Sequences (number of videos), along with the number of RGB and Depth cameras in the capture setup. For the fashion-centric comparison, our criteria include: MV for Multi-View (use of multiple synchronized cameras), Poses (different poses in the same clothing; ’multiple’ refers to video sequences with many poses), Paired (linking images of the worn outfit to garment photos) and Video (static images vs. video). In both cases, we also add Outfits (number of outfits), and Frames (number of RGB frames). In this table, wherever used, ’K’ denotes thousands (

10^{3}

) and ’M’ denotes millions (

10^{6}

). ✓ indicates available/yes, ✗ indicates not available/no/unspecified.

Figure 2: Available data for one subject in a single clothing set. (a) Subject: Frontal view, clothing layers (w & w/o jacket), styles (open & closed jacket) and template recordings, four representative RGB views, and depth images with the reconstructed point cloud. (b) Image Annotations: Foreground segmentation isolating the subject, garment segmentation for each clothing item, fitted SMPL-X body model, and labelled bounding boxes for all garments in each frame. (c) Garments, Materials & Sizing: For each garment, frontal and back views in stretched and normal states, text description of the clothing, sizing chart with corresponding measurements (in centimeters), and properties.

	shirt- blouse	pants	shorts	dress	jacket	sweater	sweatshirt	skirt	coat	jumpsuit	vest	cardigan	blazer	tights- stockings
Total	38.3	21.2	11.7	5.2	5.1	3.8	3.6	3.4	3.0	1.3	1.1	0.9	0.7	0.1
Layer	51.6	0.0	0.0	33.3	97.4	82.1	81.5	0.0	100	70.0	100	100	100	0.0
Style	7.7	0.0	0.0	0.0	86.8	39.2	55.5	0.0	59.1	10.0	62.5	28.6	60.0	0.0

Table 2: This table shows the distribution of different clothing categories across the dataset (Total). It also shows how often each clothing type shows up in multi-layered outfits (Layer) or in outfits with style variations (Style). All values are percentages.

Benchmark	Experiment	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$
Single-View	IDM-VTON [ 13 ]	$0.881$	$0.086$	$10.187$
Single-View	InsertAnything [ 72 ]	$0.927$	$0.065$	$8.192$
Multi-View (IDM-VTON)	Cross-view	$0.868$	$0.098$	$12.907$
Multi-View (IDM-VTON)	View-adaptive	$0.873$	$0.093$	$12.775$

Table 3: Quantitative results benchmarking Single-View Baselines and Multi-View Geometric Analysis using the proposed MV-Fashion’s subsets described in Sec. 4.1 .

Figure 3: Qualitative comparison showing models trained with styling-augmented data respond to fine-grained styling prompts ( a jacket, the outer wear is fully open ), producing outputs that are sometimes distinguishable from no-style prompts ( a jacket ).

Figure 4: Results on (a) Cross-View Geometric Test vs (b) View-Adaptive Try-On . The updated IDM-VTON architecture can map between the catalogue view and the person’s pose when both front and rear images of the garment are provided to the model.

Figure 5: Qualitative results of the canonical garment normal predictor

\Psi

for groups G2, G4, G5 and G6; showing input frame, ground truth (

G^{t}

), and the predicted normal image.

Group	Per-Group	Multi-Task	Multi-Task + SwinV2
G1	$4.069$ $(5.861)$	$4.361$ $(6.061)$	$3.933$ $(5.994)$
G2	$3.006$ $(2.541)$	$3.185$ $(2.619)$	$3.067$ $(2.762)$
G3	$9.951$ $(8.437)$	$6.847$ $(9.174)$	$6.643$ $(10.135)$
G4	$5.092$ $(7.377)$	$5.644$ $(8.563)$	$4.436$ $(6.537)$
G5	$12.109$ $(10.313)$	$4.295$ $(2.746)$	$4.889$ $(4.633)$
G6	$7.767$ $(7.954)$	$6.549$ $(7.234)$	$6.389$ $(6.562)$

Table 4: MAE results (cm) for the size predictor

\Phi

, reported as mean (std). We compare the three variants from Sec. 4.2 across all garment groups (defined in Sec. 3.3 ) and the average.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
instant-ngp [ 58 ]	$29.120$ $(1.555)$	$0.936$ $(0.015)$	$0.089$ $(0.022)$
nerfacto [ 73 ]	$26.535$ $(2.047)$	$0.952$ $(0.011)$	$0.052$ $(0.009)$
splatfacto [ 36 ]
4 views	$24.633$ $(2.230)$	$0.933$ $(0.015)$	$0.074$ $(0.019)$
8 views	$27.299$ $(1.850)$	$0.944$ $(0.013)$	$0.050$ $(0.012)$
16 views	$28.500$ $(1.709)$	$0.949$ $(0.013)$	$0.039$ $(0.008)$

Table 5: Comparison of image quality metrics as mean (std). We compare against other methods and provide an ablation study on the number of views for splatfacto.

Figure 6: Qualitative examples of the presented NVS methods. It shows how Nerf-based methods struggle with fine details, but Gaussian Splatting can reconstruct fine textures and the wrinkles.

Figure 7: Representative samples from MV-Fashion illustrating the range of garment typologies (single vs. multi-layer) and poses captured in the dataset.

Figure 8: Examples of exposure anomalies: (a) An overexposed frame where strong illumination causes saturation and severe loss of texture. (b) An underexposed frame where insufficient lighting leads to a near-black appearance and loss of structural details.

Figure 9: Examples of images and detected keypoints (in green rectangles) used in the intrinsics (a and b) and extrinsics (c) calibration.

Figure 10: Examples of the reconstructed point clouds visualized from different viewpoints. We apply no post-processing on the point clouds, including 3D and color smoothing. We preprocess the depth images to filter the boundary noise by masking out the pixels that have an absolute difference above 15 mm with their nearby pixels.

Figure 11: A sample image of the color calibration target before and after the color adjustments.

Figure 12: Representative segmentation outputs produced by our pipeline. From left to right: flat-catalogue garment images, corresponding flat garment masks, foreground masks, and layered on-body garment segmentation.

Figure 13: Qualitative results of our SMPL-X fitting pipeline. We display the input images, the 2D keypoints detected by RTMW-x in COCO-Wholebody format, and the final fitted SMPL-X model.

Figure 14: Visual definition of the sizing charts for each garment group (G1-G6). We annotate the measuring tape locations for all 19 measurement parts used in our dataset. Red lines indicate the distances measured.

Attribute	Layer	Categories	Encoding
Torso Closure Style	L1 & L2	n/a / fully closed / partially closed / fully open	0 / 1 / 2 / 3
Long Sleeve Style	L1 & L2	n/a / rolled up / rolled down	0 / 1 / 2
Tucking Style	L1 & L2	n/a / tucked / outside	0 / 1 / 2
Hood Style	L2 only	n/a / hood up / hood down	0 / 1 / 2

Table 6: Draping Style Attribute Encodings. n/a stands for not applicable.

Figure 15: Qwen3-VL [ 65 ] garment description protocol. The input includes a detailed prompt and both views of the garment item to guide Qwen to generate descriptive and clear caption.

Figure 16: Some examples of the styling labels MV-Fashion provides. It is a unified string of numbers that encodes several styles we defined. We also provide the garment fit for each piece of clothing.

Figure 17: Additional example of text annotations MV-Fashion provides for each outfit. This includes a cloth category, material and elasticity details, our custom styling encoding and the detailed description.

	shirt- blouse	pants	shorts	dress	jacket	sweater	sweatshirt	skirt	coat	jumpsuit	vest	cardigan	blazer	tights- stockings	Total
Total	$38.3$	$21.2$	$11.7$	$5.2$	$5.1$	$3.8$	$3.6$	$3.4$	$3.0$	$1.3$	$1.1$	$0.9$	$0.7$	$0.1$	$100.0$
Layer	$51.6$	$0.0$	$0.0$	$33.3$	$97.4$	$82.1$	$81.5$	$0.0$	$100.0$	$70.0$	$100.0$	$100.0$	$100.0$	$0.0$	$39.02$
Style	$7.7$	$0.0$	$0.0$	$0.0$	$86.8$	$39.2$	$55.5$	$0.0$	$59.1$	$10.0$	$62.5$	$28.6$	$60.0$	$0.0$	$14.13$
Slim	$12.2$	$5.0$	$0.0$	$15.4$	$7.9$	$3.6$	$14.3$	$7.7$	$0.0$	$10.0$	$0.0$	$14.3$	$0.0$	$100.0$	$8.1$
Regular	$68.3$	$55.3$	$79.1$	$59.0$	$71.1$	$64.3$	$71.4$	$65.4$	$68.2$	$70.0$	$100.0$	$42.9$	$40.0$	$0.0$	$64.1$
Loose	$19.5$	$39.6$	$20.9$	$25.6$	$21.1$	$32.1$	$14.3$	$26.9$	$31.8$	$20.0$	$0.0$	$42.9$	$60.0$	$0.0$	$24.7$

Table 7: This table shows the distribution of different clothing categories across the dataset (Total). It also shows how often each clothing type shows up in multi-layered outfits (Layer) or in outfits with style variations (Style). Additionally, we show the distribution of our fitting style labels (slim, regular or loose) and the distribution of our elasticity annotation values for each garment category. The last Total column shows the percentage of each annotation in the full dataset. All values are percentages. The percentages for layers are calculated outfit-wise, while the rest of the values are calculated garment-wise.

Figure 18: We collected additional information from each subject, like height, weight, age and gender, since these could be important for downstream tasks. They also help us evaluate and ensure a balanced distribution of our dataset.

Figure 19: Distribution of different elasticity values and fit styles across the dataset. Highly elastic clothing items are rare, but the more common ones are well represented. The fit styles are also well distributed and MV-Fashion contains a high percentage of challenging, loose clothing.

Figure 20: IDM-VTON architecture adaptation to perform Semantic Controllability and View-Adaptive Try-On tasks. For Semantic Controllability, we augment the prompt

T

by styling annotation (

\mathbf{A}

). For View-Adaptive Try-On, we introduce feature fusion before conditioning, which concatenates the IP-Adapter’s features (

\mathbf{Z}_{IP}

) and GarmentNet’s features (

\mathbf{Z}_{GN}

) of both the frontal and rear garment views.

Figure 21: The normal predictor (

\Psi

). To disentangle intrinsic garment geometry from pose deformations, the network transforms the source garment normals

G^{s}

into a canonical, unposed state

G^{t}

. Both

G^{s}

and

P^{s}

are extracted from the input frame, and the transformation is explicitly conditioned on a target pose

P^{t}

. Crucially, both pose maps (

P^{s},P^{t}

) utilize semantic body segmentation to provide spatial correspondence cues.

Figure 22: Qualitative examples of our VTON pipeline when using in-the-wild images with complex backgrounds. Our model demonstrates effective generalization to these out-of-distribution samples. (images obtained from unsplash.com)

Figure 23: Qualitative results of IDM-VTON and InsertAnything trained on the Single-View MV-Fashion benchmark. The high fidelity achieved by both models confirms the direct compatibility of MV-Fashion with existing state-of-the-art VTON datasets. The columns represent: (a) Input (Garment and Person Mask), (b) IDM-VTON Try-On Result, (c) InsertAnything Try-On Result, (d) Ground Truth.

Figure 24: Extended qualitative results for Semantic Controllability. The original experiment augments draping style in prompt when training and compare (a) no-style to (b) draping styling prompt ( e.g . outerwear is fully open for both rows) at test time. Additionally, we test the second experiment to train with fitting style and compare (c) no-style to (b) fitting styling prompt ( e.g . regular fit for row 1 and slim fit for row 2) at test time.

Figure 25: More qualitative results of IDM-VTON on (a) Cross-View Geometric Test vs (b) View-Adaptive Try-On . The updated IDM-VTON architecture can map between the catalogue view and the person’s pose when both frontal and rear images of the garment are provided to the model.

Measurement	Per-Group	Multi-Task	Multi-Task + SwinV2
Neck	$2.400$ $(1.976)$	$2.581$ $(2.163)$	$2.246$ $(1.852)$
Chest	$5.170$ $(4.053)$	$6.370$ $(4.423)$	$5.339$ $(3.946)$
Waist	$4.265$ $(4.114)$	$4.516$ $(4.344)$	$4.581$ $(4.490)$
Bottom	$3.553$ $(2.921)$	$4.347$ $(2.652)$	$3.659$ $(3.118)$
Sleeve	$7.903$ $(14.074)$	$8.022$ $(14.391)$	$7.058$ $(14.679)$
Bicep	$1.672$ $(1.177)$	$1.838$ $(1.418)$	$1.620$ $(1.096)$

Table 8: Detailed breakdown of Tab. 4 for group G1.

Measurement	Per-Group	Multi-Task	Multi-Task + SwinV2
Bottom Waist	$4.177$ $(2.573)$	$4.266$ $(2.607)$	$3.572$ $(2.696)$
Bottom Hip	$3.223$ $(2.249)$	$3.609$ $(2.263)$	$2.993$ $(2.291)$
Thigh	$1.697$ $(1.302)$	$2.027$ $(1.515)$	$1.767$ $(1.316)$
Leg Cuff	$2.459$ $(2.415)$	$2.566$ $(2.669)$	$1.991$ $(2.171)$
Front Crotch	$2.561$ $(1.854)$	$2.803$ $(1.977)$	$1.946$ $(1.709)$
Back Crotch	$2.552$ $(2.404)$	$2.603$ $(2.402)$	$2.784$ $(2.243)$

Table 9: Detailed breakdown of Tab. 4 for group G2.

Measurement	Per-Group	Multi-Task	Multi-Task + SwinV2
Bottom Waist	$6.325$ $(4.495)$	$3.938$ $(3.331)$	$4.367$ $(2.685)$
Bottom Hip	$8.126$ $(4.236)$	$3.894$ $(2.101)$	$3.760$ $(2.766)$
Bottom Bottom	$17.300$ $(12.239)$	$15.506$ $(14.732)$	$14.483$ $(17.524)$
Full Length	$8.053$ $(5.055)$	$4.049$ $(2.694)$	$3.962$ $(3.377)$
Average	$9.951$ $(8.437)$	$6.847$ $(9.174)$	$6.643$ $(10.135)$

Table 10: Detailed breakdown of Tab. 4 for group G3.

Measurement	Per-Group	Multi-Task	Multi-Task + SwinV2
Neck	$1.768$ $(1.277)$	$2.444$ $(2.284)$	$1.815$ $(1.955)$
Chest	$6.279$ $(4.443)$	$7.493$ $(5.830)$	$5.996$ $(4.814)$
Waist	$4.656$ $(3.858)$	$4.946$ $(5.611)$	$3.721$ $(3.169)$
Bottom	$6.131$ $(5.767)$	$6.922$ $(6.804)$	$5.171$ $(5.246)$
Sleeve	$7.784$ $(6.336)$	$5.717$ $(4.442)$	$5.842$ $(4.986)$
Bicep	$2.319$ $(1.907)$	$2.116$ $(1.439)$	$1.619$ $(1.300)$

Table 11: Detailed breakdown of Tab. 4 for group G4.

Measurement	Per-Group	Multi-Task	Multi-Task + SwinV2
Neck	$8.058$ $(2.290)$	$3.641$ $(2.401)$	$1.132$ $(0.590)$
Chest	$10.098$ $(1.426)$	$4.768$ $(2.114)$	$5.174$ $(1.231)$
Waist	$10.615$ $(1.759)$	$2.701$ $(1.207)$	$2.721$ $(1.042)$
Bottom	$11.753$ $(2.834)$	$3.985$ $(2.504)$	$3.945$ $(2.169)$
Armhole	$2.083$ $(1.326)$	$6.951$ $(1.245)$	$10.211$ $(2.226)$
Shoulder	$9.791$ $(1.785)$	$6.528$ $(3.385)$	$6.931$ $(3.096)$

Table 12: Detailed breakdown of Tab. 4 for group G5. Note: Sleeve , Bicep , and Sleeve Cuff measurement parts are missing because all instances in the test set are sleeveless.

Measurement	Per-Group	Multi-Task	Multi-Task + SwinV2
Neck	$4.285$ $(2.699)$	$2.944$ $(2.135)$	$2.995$ $(2.366)$
Chest	$8.361$ $(10.686)$	$6.902$ $(7.048)$	$6.050$ $(5.117)$
Waist	$7.555$ $(7.777)$	$4.553$ $(3.660)$	$3.700$ $(3.120)$
Bottom	$8.080$ $(7.073)$	$4.268$ $(3.630)$	$3.787$ $(2.954)$
Sleeve	$6.575$ $(4.483)$	$4.183$ $(2.600)$	$6.146$ $(4.889)$
Bicep	$3.446$ $(2.441)$	$2.957$ $(2.528)$	$3.781$ $(2.119)$

Table 13: Detailed breakdown of Tab. 4 for group G6.

Group	Per-Group	Multi-Task	Multi-Task + SwinV2	End-to-End
G1	$4.069$ $(5.861)$	$4.361$ $(6.061)$	$3.933$ $(5.994)$	$5.316$ $(6.672)$
G2	$3.006$ $(2.541)$	$3.185$ $(2.619)$	$3.067$ $(2.762)$	$3.965$ $(3.789)$
G3	$9.951$ $(8.437)$	$6.847$ $(9.174)$	$6.643$ $(10.135)$	$10.565$ $(13.070)$
G4	$5.092$ $(7.377)$	$5.644$ $(8.563)$	$4.436$ $(6.537)$	$5.924$ $(6.705)$
G5	$12.109$ $(10.313)$	$4.295$ $(2.746)$	$4.889$ $(4.633)$	$6.347$ $(5.192)$
G6	$7.767$ $(7.954)$	$6.549$ $(7.234)$	$6.389$ $(6.562)$	$11.076$ $(11.976)$

Table 14: Quantitative results for the size regressor (

\Phi

). We report MAE in cm as mean (std). This table expands Tab. 4 with the End-to-End baseline, which regresses size directly from the input image. The comparison assesses the performance with and without relying on ground truth canonical garment normals (

G^{t}

). Groups and model variants are defined in Sec. 3.3 and Sec. 4.2 of the main paper, respectively.

Figure 26: Qualitative results of the garment normal predictor (

\Psi

) across all garment groups (G1-G6). From top to bottom: the input RGB frame, the source garment normal map (

G^{s}

), the target ground truth canonical normal map (

G^{t}

), and the predicted canonical normals. The results demonstrate the model’s ability to disentangle pose from garment geometry, effectively recovering the intrinsic shape even in cases of occlusion (e.g., G1) or complex deformations (e.g., G4).

Figure 27: Additional examples of the three NVS methods tested. Instant-NGP produces blurry results, while nerfacto is sharper, but it has some distinct artifacts. splatfacto on the other hand provides superior quality.

Figure 28: Qualitative results of the ablation over the number of views used for splatfacto when using only Bolt cameras. Because of the limited number of available cameras there are many artifacts.

Figure 29: Qualitative results of the ablation over the number of views used for splatfacto . It shows that using very few views produces artifacts, but from 16 views it starts to improve and when using all views the results are almost indistinguishable from the ground truth.

Views	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
4	$24.089(1.806)$	$0.952(0.012)$	$0.073(0.016)$
6	$25.831(1.777)$	$0.956(0.012)$	$0.061(0.013)$

Table 15: Image quality metrics when using only Bolt cameras with splatfacto . Note that these values are not directly comparable with Tab. 5 due to the resolution difference. Also, because of the limited number of cameras, these values are relatively low.

Figure 30: Qualitative results of the dynamic novel view synthesis experiments. The results show slight blurring around the hands, but otherwise prvode good quality outputs. The frames shown are 6 consecutive frames from a 20 frame sequence.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
CEM-4DGS	$24.395$	$0.875$	$0.151$

Table 16: Image quality metrics of the 4DGS reconstruction. As expected, slightly lower than static reconstruction, but still good quality. Values shown are means across all frames of the sequence.

MV-Fashion: Towards Enabling Virtual Try-On and Size Estimation with Multi-View Paired Data

Abstract

Figures & tables

Virtual Try-On

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

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

Layering Virtual Try-On

Abstract

Virtual Try-On

Explore similar work

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

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

Layering Virtual Try-On