BridgeDiff: Bridging Human Observations and Flat-Garment Synthesis for Virtual Try-Off

Shuang Liu; Ao Yu; Linkang Cheng; Xiwen Huang; Li Zhao; Junhui Liu; Zhiting Lin; Yu Liu

cs.CVMar 10, 2026

BridgeDiff: Bridging Human Observations and Flat-Garment Synthesis for Virtual Try-Off

Authors: Shuang Liu, Ao Yu, Linkang Cheng, Xiwen Huang, Li Zhao, Junhui Liu, Zhiting Lin, Yu Liu

Organizations: School of Integrated Circuits, Anhui University, Hefei, China · School of Astronautics, Northwestern Polytechnical University, Xi’an,China · Anhui Provincial High-performance Integrated Circuit Engineering Research Center

Abstract

Virtual try-off (VTOFF) aims to recover canonical flat-garment representations from images of dressed persons for standardized display and downstream virtual try-on. Prior methods often treat VTOFF as direct image translation driven by local masks or text-only prompts, overlooking the gap between on-body appearances and flat layouts. This gap frequently leads to inconsistent completion in unobserved regions and unstable garment structure. We propose BridgeDiff, a diffusion-based framework that explicitly bridges human-centric observations and flat-garment synthesis through two complementary components. First, the Garment Condition Bridge Module (GCBM) builds a garment-cue representation that captures global appearance and semantic identity, enabling robust inference of continuous details under partial visibility. Second, the Flat Structure Constraint Module (FSCM) injects explicit flat-garment structural priors via Flat-Constraint Attention (FC-Attention) at selected denoising stages, improving structural stability beyond text-only conditioning. Extensive experiments on standard VTOFF benchmarks show that BridgeDiff achieves state-of-the-art performance, producing higher-quality flat-garment reconstructions while preserving fine-grained appearance and structural integrity.

Figures & tables

Figure 1: Existing methods often suffer from visual discontinuity and structural instability when relying solely on textual conditioning or weak local constraints, especially under occlusions or partial observations. In contrast, BridgeDiff bridges dressed-person observations and canonical flat-garment representations via garment cues representation and explicit flat structure guidance.

Figure 2: Overview of the proposed GCBM. Rather than directly mapping dressed-person observations to flat garment images, GCBM aggregates multiple visual information into garment cues representation, capturing the global appearance and identity of the target garment to support visually continuous flat-garment synthesis.

Figure 3: Overview of the proposed Flat Structure Constraint for Conditional Diffusion architecture. The framework consists of a trainable model UNet and a largely frozen denoising UNet. To explicitly enforce flat garment layouts, a flat structure constraint module (FSCM) is integrated into the denoising UNet, ensuring stable layout generation without compromising appearance fidelity.

	All						Upper-Body
Method	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
Any2AnyTryOn [ 9 ]	17.25	7.46	12.80	74.43	36.68	24.12	15.29	2.14	12.97	74.48	37.74	23.38
MGT ∗ [ 37 ]	12.62	7.56	12.72	75.35	35.01	24.73	19.18	6.65	12.27	74.38	40.24	25.42
BridgeDiff (Ours)	10.92	3.86	16.69	80.23	23.10	20.82	17.60	4.96	15.90	79.07	27.52	22.36
	Lower-Body						Dresses
Method	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$

Table 1: Quantitative results on the DressCode dataset over the full test set and category-specific subsets.

\uparrow

indicates higher is better,

\downarrow

indicates lower is better. Bold denotes the best result and underline denotes the second-best result.

*

denotes results obtained using a single unified model trained jointly on all three garment categories.

Figure 4: Qualitative comparisons on the DressCode dataset. Red circles highlight differences in local regions across different methods. Unmarked examples indicate cases where the overall garment structure or color appearance differs from the reference. Zooming in provides a clearer view of these differences.

Method	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
TryOffDiff [ 36 ]	22.54	8.58	11.77	71.83	42.27	24.87
TryOffAnyone [ 40 ]	12.59	2.77	12.51	71.90	34.47	22.68
Any2AnyTryOn [ 9 ]	12.15	1.96	12.86	74.29	35.72	22.80
Voost [ 16 ]	10.06	2.48	-	-	-	-
UniFit [ 46 ]	12.58	-	-	77.50	28.10	20.20
BridgeDiff (Ours)	9.08	1.53	15.00	77.42	24.38	18.69

Table 2: Quantitative results on the VITON-HD dataset over the full test set and category-specific subsets.

\uparrow

indicates higher is better,

\downarrow

indicates lower is better. Bold denotes the best result and underline denotes the second-best result. “–” indicates the metric is not reported.

\dagger

denotes cross-dataset evaluation where the model is trained on DressCode and evaluated on VITON-HD.

Figure 5: Qualitative comparisons on the VITON-HD dataset. Red circles highlight differences in local regions across different methods. Unmarked examples indicate cases where the overall garment structure or color appearance differs from the reference. Zooming in provides a clearer view of these differences.

Effectiveness of the GCBM
Method	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
w/o GCBM	9.37	1.51	15.05	76.91	24.93	19.34
w/ GCBM (Ours)	9.08	1.53	15.00	77.42	24.38	18.69
Effectiveness of the FSCM
M1 (w/o $T_{c}$ )	9.26	1.89	15.10	76.93	24.14	18.71
M2 (w/o $F_{c}^{\prime}$ )	9.46	1.62	15.22	77.11	24.11	18.96

Table 3: Quantitative comparison of GCBM and FSCM on VITON-HD. The upper part reports the ablation results of GCBM, while the lower part presents the comparison among different FSCM variants (M1–M3).

Figure 6: Qualitative comparisons on the VITON-HD dataset. Left: GCBM qualitative analysis. Right: FSCM qualitative analysis. Red circles highlight differences in local regions across different methods. Zooming in provides a clearer view of these differences.

Figure 7: Heatmap visualization at the 50th denoising step on the VITON-HD dataset. We compare the full model with variants without GCBM or FSCM. Red circles highlight differences in local regions across different methods.

Figure 8: Summary of the user study results. (a) VITON-HD. (b) DressCode. Within each dataset, the left panel shows the average perceptual scores of different methods across three dimensions as bar charts, while the right panel presents the overall user preference distribution as pie charts.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition
$X_{m}$	Model image
$X_{w}$	Warped garment
$X_{c}$	Target garment image
$F_{w}$	Encoded feature of warped garment
$F_{m}$	Encoded feature of model image (stage 1)
$F_{c}$	Output of garment condition bridge module (GCBM)

Appendix

Table 4: Notation Definition

Method	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
Down 1	9.42	1.86	15.00	77.49	24.18	18.82
Down 2	9.32	1.88	15.03	77.24	24.27	18.77
Mid	9.25	1.84	15.07	77.36	24.22	18.77
Up 1	9.42	1.81	15.03	77.48	24.20	18.73
Up 2	9.03	1.67	15.05	77.00	24.42	18.71
Up 3	9.35	1.85	14.92	77.06	24.37	18.71

Appendix

Table 5: Ablation study on FSCM insertion positions on the VITON-HD dataset.

\uparrow

indicates higher is better,

\downarrow

indicates lower is better. Bold denotes the best result.

Figure 9: Qualitative comparisons of FSCM Insertion Position on the VITON-HD dataset. Red circles highlight differences in local regions across different settings. Zooming in provides a clearer view of these differences.

Comparison with UniFit’s Spatial Attention Focusing Loss
Method	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
w/ Loss (UniFit)	9.78	1.99	14.92	77.37	24.67	18.66
w/o Loss (Ours)	9.08	1.53	15.00	77.42	24.38	18.69
Impact of Textual Descriptions
w/o Text	9.06	1.57	15.00	77.07	24.42	18.76
w/ Text (Ours)	9.08	1.53	15.00	77.42	24.38	18.69

Appendix

Table 6: Ablation study results on the VITON-HD dataset.

\uparrow

indicates higher is better,

\downarrow

indicates lower is better. Bold denotes the best result. The upper part reports the comparison with UniFit’s spatial attention focusing loss, while the lower part presents the impact of textual descriptions.

Panel of Figure 10 — Figure 10: Qualitative comparisons on the VITON-HD dataset. Red circles highlight differences in local regions across different settings. Zooming in provides a clearer view of these differences.

Figure 11: Qualitative comparisons on the VITON-HD dataset. Upper: DREAM qualitative analysis. Lower: CFG qualitative analysis. Red circles highlight differences in local regions across different methods. Zooming in provides a clearer view of these differences.

$p$	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
0.0	11.92	3.58	14.99	77.58	25.09	20.51
0.5	9.60	2.04	14.96	77.41	24.60	19.14
1.0 (Ours)	9.08	1.53	15.00	77.42	24.38	18.69
1.5	9.13	1.58	14.89	76.86	24.51	18.66
2.0	9.42	1.84	14.85	76.66	24.44	18.75
$\infty$	9.15	1.63	14.87	77.28	24.64	18.66

Appendix

Table 7: Ablation results of different

p

in DREAM on the VITON-HD dataset.

p=\infty

means DREAM is disabled.

\uparrow

indicates higher is better,

\downarrow

indicates lower is better. Bold denotes the best result.

Figure 12: Overview of the textual description generation pipeline. Left: multiple MLLMs are employed to discover and filter visually grounded garment attributes. Right: the finalized structured attributes are used to guide Qwen3-VL for controllable garment description generation.

Figure 13: An example of the user study questionnaire for a single image set. Participants are presented with multiple generated results for the same model image and are asked to rate each result based on visual realism, garment structure consistency, occluded region continuity, and overall preference.

Figure 14: Additional qualitative comparisons on the DressCode dataset. Red circles highlight differences in local regions across different methods. Unmarked examples indicate cases where the overall garment structure or color appearance differs from the reference. Zooming in provides a clearer view of these differences.

Figure 15: Additional qualitative comparisons on the DressCode dataset. Red circles highlight differences in local regions across different methods. Unmarked examples indicate cases where the overall garment structure or color appearance differs from the reference. Zooming in provides a clearer view of these differences.

Figure 16: Additional qualitative comparisons on the VITON-HD dataset. Red circles highlight differences in local regions across different methods. Unmarked examples indicate cases where the overall garment structure or color appearance differs from the reference. Zooming in provides a clearer view of these differences.

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

What Matters in Virtual Try-Off? Dual-UNet Diffusion Model For Garment Reconstruction
Apr 9, 2026Loc-Phat Truong, Meysam Madadi, Sergio EscaleraDiffusion Models Diffusion-Based Image Generation
Computer Vision Center, Spain · Universitat de Barcelona, Spain
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
Delving into Asymmetric Information Dynamics for High-Fidelity Virtual Try-On
Sep 22, 2026Zishu Qin, Zhiyu Jin, Pipei Huang +1Virtual Try-On Conditional Diffusion Models
Alibaba Group

Apr 9, 2026J/K move · Enter open · S save

What Matters in Virtual Try-Off? Dual-UNet Diffusion Model For Garment Reconstruction

Loc-Phat Truong, Meysam Madadi, Sergio Escalera

Computer Vision Center, Spain · Universitat de Barcelona, Spain

Diffusion Models Diffusion-Based Image Generation

Open paper AskPDF arXiv

Virtual Try-On (VTON) has seen rapid advancements, providing a strong foundation for generative fashion tasks. However, the inverse problem, Virtual Try-Off (VTOFF)-aimed at reconstructing the canonical garment from a draped-on image-remains a less understood domain, distinct from the heavily researched field of VTON. In this work, we seek to establish a robust architectural foundation for VTOFF by studying and adapting various diffusion-based strategies from VTON and general Latent Diffusion Models (LDMs). We focus our investigation on the Dual-UNet Diffusion Model architecture and analyze three axes of design: (i) Generation Backbone: comparing Stable Diffusion variants; (ii) Conditioning: ablating different mask designs, masked/unmasked inputs for image conditioning, and the utility of high-level semantic features; and (iii) Losses and Training Strategies: evaluating the impact of the auxiliary attention-based loss, perceptual objectives and multi-stage curriculum schedules. Extensive experiments reveal trade-offs across various configuration options. Evaluated on VITON-HD and DressCode datasets, our framework achieves state-of-the-art performance with a drop of 9.5% on the primary metric DISTS and competitive performance on LPIPS, FID, KID, and SSIM, providing both stronger baselines and insights to guide future Virtual Try-Off research.

Abstract

Virtual Try-On

Explore similar work

What Matters in Virtual Try-Off? Dual-UNet Diffusion Model For Garment Reconstruction

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

Delving into Asymmetric Information Dynamics for High-Fidelity Virtual Try-On