Paper deep dive
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
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/13/2026, 1:00:37 AM
Summary
BridgeDiff is a diffusion-based framework for Virtual Try-Off (VTOFF) that reconstructs canonical flat-garment representations from images of dressed persons. It addresses visual discontinuity and structural instability by introducing two components: the Garment Condition Bridge Module (GCBM) for global garment-cue representation and the Flat Structure Constraint Module (FSCM) for injecting explicit structural priors via Flat-Constraint Attention.
Entities (6)
Relation Signals (4)
BridgeDiff → includes → GCBM
confidence 100% · We propose BridgeDiff... through two complementary components... Garment Condition Bridge Module (GCBM)
BridgeDiff → includes → FSCM
confidence 100% · We propose BridgeDiff... through two complementary components... Flat Structure Constraint Module (FSCM)
BridgeDiff → performs → VTOFF
confidence 98% · BridgeDiff, a structure-aware diffusion framework for the VTOFF task
BridgeDiff → evaluatedon → DressCode
confidence 95% · Extensive experiments on standard VTOFF benchmarks... Quantitative results on the DressCode dataset
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2603.09236v1
- Canonical: https://arxiv.org/abs/2603.09236v1
Trouble viewing inline? Open PDF directly →
Full Text
79,167 characters extracted from source content.
Expand or collapse full text
BridgeDiff: Bridging Human Observations and Flat-Garment Synthesis for Virtual Try-Off Shuang Liu 1 , Ao Yu 1 , Linkang Cheng 1 , Xiwen Huang 1 , Li Zhao 1 , Junhui Liu 3 , Zhiting Lin 1,2 , and Yu Liu 1,2⋆ 1 School of Integrated Circuits, Anhui University, Hefei, China 2 Anhui Provincial High-performance Integrated Circuit Engineering Research Center 3 School of Astronautics, Northwestern Polytechnical University, Xi’an,China 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 di- rect image translation driven by local masks or text-only prompts, over- looking 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 rep- resentation that captures global appearance and semantic identity, en- abling robust inference of continuous details under partial visibility. Sec- ond, 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 bench- marks show that BridgeDiff achieves state-of-the-art performance, pro- ducing higher-quality flat-garment reconstructions while preserving fine- grained appearance and structural integrity. Keywords: Virtual Try-Off· Fashion· Diffusion Model 1 Introduction In the rapidly evolving landscape of e-commerce, particularly within the fash- ion industry, providing realistic, flexible, and scalable visual representations of garments has become increasingly important for enhancing user experience and driving consumer engagement. While virtual try-on (VTON) [3–5,9,13,15,16,28] techniques enable customers to visualize how a garment appears when worn by a person, they typically rely on standardized catalog images that are costly to produce. To address this limitation, the recently proposed virtual try-off (VTOFF) [9,16,17,36,37,40,46] task aims to reconstruct clean, canonical flat- garment representations directly from images of dressed individuals. By recover- ing garments in a standardized, catalog-style format, VTOFF offers significant ⋆ Corresponding authors arXiv:2603.09236v1 [cs.CV] 10 Mar 2026 2S. Liu et al. value for product presentation, visual retrieval, and downstream applications (e.g., person-to-person try-on). However, accurately reconstructing flat-garments from on-body observations remains challenging due to severe occlusions, pose variations, and the inherent distribution gap between human-centric images and canonical flat-garment layouts. Effectively bridging this gap while ensuring visual continuity of garment details and maintaining stable garment structure under limited or partially observable conditions is therefore a key factor for advancing VTOFF performance. Existing studies on VTOFF can be broadly categorized into two lines of research. The first line focuses on standalone VTOFF modeling, formulating the task as reconstructing clean and canonical flat-garment images directly from photos of dressed person. Representative works include TryOffDiff [36], which pioneers diffusion-based garment reconstruction, MGT [37], which intro- duces category-specific embeddings to support multiple garment types, and Try- OffAnyone [40], which leverages explicit garment masks and lightweight fine- tuning to balance efficiency and quality. The second line of research aims to unify VTON and VTOFF within a single framework to improve generality and scalability. Methods including Any2AnyTryOn [9] and Voost [16] adopt Diffu- sion Transformer architectures with task-aware conditioning, while OMFA [17] and UniFit [46] further enhance human–garment disentanglement and seman- tic alignment under diverse poses and multimodal instructions. Despite these advances, most existing VTOFF methods still formulate the task as a direct mapping from dressed-person images to flat-garment images, typically relying on local mask constraints or coarse textual cues. Such designs struggle to handle partially observable regions, often leading to discontinuous or implausible gar- ment details under occlusions and pose variations. Moreover, although unified multi-task frameworks improve flexibility, their reliance on text-dominant con- ditioning makes it difficult to consistently enforce stable flat-garment structures during generation. These limitations highlight the need for explicitly modeling garment cues representation and introducing structured constraints to ensure both visual continuity and structural stability in VTOFF. Motivated by these observations, we propose BridgeDiff, a structure-aware diffusion framework for the VTOFF task, aiming to generate visually contin- uous and structurally stable flat-garments from human-centric observation im- ages. As illustrated in Fig. 1, existing VTOFF methods often suffer from two key limitations. First, under occlusions or partial observations, the absence of garment cues representation makes it difficult to infer visually continuous details for unseen regions. Second, relying solely on textual descriptions or weak local constraints is insufficient to enforce stable flat-garment geometry during genera- tion. To address these challenges, BridgeDiff explicitly bridges the gap between dressed-person observations and canonical flat garment representations by intro- ducing garment cues representation and explicit flat structure guidance within the diffusion process. Specifically, we design a garment condition bridge module (GCBM) to aggregate global garment appearance from dressed-person images into garment cues representation, facilitating continuous inference of garment de- BridgeDiff for Virtual Try-Off3 Text Prompt Reconstruct a top with waist wrap, fit fitted, hem peplum, neckline v-neck, sleeve length long, cloth length upper body Model Images Diffusion Model A flat-lay top FSCM Text1 Prompt Reconstruct a top with waist wrap, fit fitted, hem peplum, neckline v-neck, sleeve length long, cloth length upper body Diffusion Model √ Weak controllability × Model Images Strong controllability (b1) Existing Methods (b2) FSCM (Ours) Text2 Prompt × Poor Flat-lay √ Natural Flat-lay Reconstruct a top with waist regular, fit loose, hem straight, neckline round, sleeve length short, cloth length short (a1) Existing Methods Diffusion Model × Discontinuous × With Artifacts Reconstruct a top with waist regular, fit loose, hem straight, neckline round, sleeve length short, cloth length short Diffusion Model Continuous Artifacts-free √ √ GCBM (a2) GCBM (Ours) Text Prompt Text Prompt Fig. 1: Existing methods often suffer from visual discontinuity and structural insta- bility 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 representa- tion and explicit flat structure guidance. tails beyond visible regions. In addition, we introduce a flat structure constraint module (FSCM), which injects explicit flat-garment structural information into a specific stage of the diffusion denoising UNet, further improving the structural stability of the generated results. Our contributions are summarized as follows: – We propose a garment condition bridge module (GCBM) that constructs garment cues representation, enabling stable modeling of global garment appearance and semantic identity, and supporting visually continuous detail generation under limited conditioning in VTOFF scenarios. – We introduce flat structure constraint module (FSCM) that injects explicit flat-garment structural information into the diffusion denoising process via structure-aware attention, significantly improving structural stability and geometric plausibility. – We conduct comprehensive experiments on two public VTOFF benchmarks, together with a user study, demonstrating the effectiveness of BridgeDiff in terms of both quantitative performance and visual quality. 2 Related Work Virtual Try-On. Image-based virtual try-on (VTON) synthesizes a person wearing a target garment while preserving pose, body shape, and identity. Early VTON methods are largely GAN-based [3, 7, 15, 38], which often suffer from training instability and limited fidelity on fine garment details. Diffusion mod- els [10,29,30,32] have recently become the dominant paradigm for VTON [4,5,8, 20,28,42] due to their strong generation quality. Early diffusion-based approaches commonly follow warping-based pipelines [8, 20], which can introduce artifacts under imperfect geometric alignment. Warping-free diffusion methods [4,41,48] and designs with garment encoders or dual UNet architectures [13, 28, 31] fur- ther improve clothing fidelity. However, most VTON work focuses on the forward synthesis setting, while the inverse setting remains less studied. 4S. Liu et al. Virtual Try-Off. It aims to reconstruct clean garment representations from images of dressed persons. TryOffDiff [36] initiates this direction with diffusion- based garment generation, and MGT [37] extends it with category-aware em- beddings for multi-garment modeling. TryOffAnyone [40] incorporates garment masks and parameter-efficient tuning to balance quality and efficiency. Recent works further move toward unifying VTON and VTOFF. Any2AnyTryOn [9] and Voost [16] adopt DiT backbones [24] with task conditioning for scalable multi-task learning, while OMFA [17] and UniFit [46] explore bidirectional gener- ation and semantic alignment via richer conditioning. Despite this progress, most VTOFF methods regress flat garments directly from dressed-person images using category text or local masks, while neglecting the distribution gap between oc- cluded, pose-dependent observations and canonical flat-garment structure. This often yields discontinuous details and unstable geometry. 3 Methodology 3.1 Garment Condition Bridge Module A fundamental challenge in VTOFF lies in the distribution gap between human- centric observation images and canonical flat-garment synthesis space. In conven- tional VTON tasks, this gap can be largely mitigated by leveraging rich human- related conditions (e.g., pose information, human parsing maps, and garment masks), which help models focus on the target garment regions. Image Encoder B Transformer Encoder Garment Cues Position Embedding Position Embedding B B ** B ** B ** B * *** Learnable TokenLearnable Token Gaussian NoiseGaussian Noise Multiply Fig. 2: Overview of the proposed GCBM. Rather than directly mapping dressed-person observations to flat gar- ment images, GCBM aggregates multi- ple visual information into garment cues representation, capturing the global ap- pearance and identity of the target gar- ment to support visually continuous flat-garment synthesis. However, in VTOFF scenarios, many of these conditions are unavailable or unreliable, making it difficult for the model to accurately recover visually con- tinuous details in occluded regions un- der sparse or partial constraints. To ad- dress this challenge, we propose explicitly constructing garment cues representation that captures the global appearance and semantic identity of the target garment. Motivated by the hierarchical repre- sentation modeling strategy adopted in DALL·E 2 [26], we introduce the GCBM to construct garment cues representation for VTOFF. As illustrated in Fig. 2, we first apply a cloth-agnostic mask to the reference model image X m to obtain a warped garment representation X w . Then, the warped garment image X w , the reference model image X m , and the corresponding target garment image X c are independently encoded by a shared BridgeDiff for Virtual Try-Off5 image encoder, producing feature repre- sentations F w , F m , and F c , respectively. To align with the diffusion timestep, the target garment feature F c is perturbed through a forward noising process 4 , yielding a noisy feature representation de- noted as ˆ F c . Subsequently, The feature tokens F w , F m , and ˆ F c are concatenated along the token dimension and fed into the GCBM. The internal architecture of GCBM follows MetaFormer [44] design, enabling flexible token mixing across heterogeneous feature sources. To distinguish different conditional features, we use learnable positional encodings for each type, enabling effective differentiation of different conditional feature during aggregation. The training objective of the GCBM, parameterized by θ, is defined as follows: L GCBM = E F c ,F w ,F m ,ε,t h F c − f θ ( ˆ F c ,F w ,F m ,t) 2 2 i , (1) where t denotes the diffusion timestep. At inference time, GCBM starts from a pure noise initialization and iteratively denoises it to progressively refine the target garment cues representation. At each diffusion step t, the model conditions on the model image feature F m and the warped garment representation F w , and predicts the clean garment cues representation as: F c = f θ (ε,F w ,F m ,t),(2) where ε denotes the gaussian noise. After completing the iterative denoising process 5 , the model obtains approximately garment cues. This cues serves as a semantic bridge between human-centric observations and the canonical flat garment layout, and is subsequently used as a conditioning signal for downstream diffusion-based synthesis. 3.2 Flat Structure Constraint for Conditional Diffusion Overview. Our goal is to generate structurally stable flat garment images while preserving fine-grained appearance details. To this end, we propose a Flat Struc- ture Constraint for Conditional Diffusion framework to explicitly enforce the flat layout of garments. As illustrated in Fig. 3, the proposed framework consists of three main components: a model UNet, a denoising UNet, and a FSCM. The parameters of the denoising UNet are largely frozen, and FSCM is incorporated into it to maintain the structural stability of the generated flat garments. Mean- while, the model UNet serves as a conditional feature extractor, which captures fine-grained garment details from the input model image. Model Unet. Unlike conventional diffusion frameworks that rely solely on image encoders for conditional guidance, the model UNet directly processes the input 4 In our experiments, we adopt the forward noising process defined in DDPM [10]. The implementation is based on https://github.com/hojonathanho/diffusion. 5 At inference time, we adopt the DDIM [33] reverse process for efficient sampling, following the implementation in https://github.com/ermongroup/ddim. 6S. Liu et al. Image Encoder Projection VAE Encoder Model Image ... ... VAE Encoder Garment Cues Class Projection Noise Image Timestep ++ ... Caption Reconstruct a knit with hem straight, neckline round, waist regular, sleeve length three quarter, cloth length short, fit regular. Text Encoder Text Encoder Model Unet Denoising Unet Vanilla Self-Attention Hybrid Cross-Attention ...... Resnet Block Hybrid Attention Hybrid Cross-Attention Garment Cues Text Encoder <A flat-lay knit> Flat Constraint Attention Flat Structure Constraint Module LinearLinear VAE Decoder Target Image Cross-AttentionCross-Attention ++ Cross-AttentionCross-Attention + Caption Trainable Frozen Add Multiply + Trainable Frozen Add Multiply + A model is wearing a knit with hem straight, neckline round, waist regular, sleeve length three quarter, cloth length short, fit regular. Flat Constraint Attention LinearLinear LinearLinear LinearLinear Multi - Head Cross - Attention Flat Constraint Attention Linear Linear Linear Multi - Head Cross - Attention Fig. 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 genera- tion without compromising appearance fidelity. model image and its textual description to extract intermediate garment features. During both training and inference, the latent variables are kept noise-free and a single forward pass is performed at diffusion timestep T = 0, following the design adopted in IMAGDressing-v1 [28] and IDM-VTON [4]. The extracted intermediate features are then used as conditional guidance for the denoising UNet. Specifically, given a model image X m ∈ R 3×H×W , a frozen Variational Autoencoder (VAE) encoder and an image encoder are applied to obtain the latent representation Z m ∈ R 4× H 8 × W 8 and the image embedding I m ∈ R 1024 , respectively. The corresponding textual description is encoded by a text encoder to produce T m ∈ R 77×768 . The image embedding I m ∈ R 1024 is projected to I ′ m ∈ R 768 , and the hybrid cross-attention at the i-th layer is defined as: Z (i) hc = Softmax Q (i) K ⊤ T √ d V T + λ Softmax Q (i) K I ⊤ √ d ! V I ,(3) BridgeDiff for Virtual Try-Off7 Table 1: Quantitative results on the DressCode dataset over the full test set and category-specific subsets. ↑ indicates higher is better, ↓ 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. AllUpper-Body MethodFID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ FID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ Any2AnyTryOn [9] 17.25 7.4612.8074.43 36.68 24.1215.29 2.14 12.9774.48 37.7423.38 MGT ∗ [37]12.627.56 12.72 75.3535.0124.73 19.18 6.65 12.27 74.3840.24 25.42 BridgeDiff (Ours)10.923.8616.6980.2323.1020.8217.604.9615.9079.0727.5222.36 Lower-BodyDresses MethodFID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ FID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ Any2AnyTryOn 49.04 22.37 12.51 74.47 37.55 25.51 28.78 14.14 12.91 74.35 34.74 23.47 MGT ∗ 22.156.4812.5475.2134.5824.6320.356.0913.4176.4630.2224.13 BridgeDiff (Ours)19.865.1316.7180.9922.7121.0816.374.7217.5280.6519.0719.03 where λ is a hyperparameter used to regulate the contribution of image-conditioned attention. Specifically, the query, key, and value matrices are defined as Q (i) = Z (i) m W q , K t = T m W k , and V t = T m W v for the text branch, while K e = I ′ m W k and V e = I ′ m W ′ v correspond to the image branch. The projection matrices W ′ k and W ′ v are initialized from W k and W v , respectively, and remain learnable dur- ing training. Denoising UNet. In the denoising UNet, intermediate features from the model UNet, denoted as M o ∈ R N×C , are used as additional condi- tioning and injected into all self-attention layers of the denoising UNet. Given a cloth-agnostic mask M c ∈ R N×1 , the masked output for an intermediate layer is defined as: M ′ o = Interpolate(M c )⊗ M o ,(4) where ⊗ denotes the element-wise (Hadamard) product, and Interpolate(·) de- notes the mask to match the spatial resolution of M o . The resulting masked features M ′ o ∈ R N×C are then injected as additional conditioning inputs into the corresponding self-attention layers of the denoising UNet. Then, we replace all vanilla self-attention layers with hybrid attention layers to incorporate condi- tional features. Given a target garment image X c ∈ R 3×H×W , we first encode it using a frozen VAE encoder to obtain the latent representation Z c ∈ R 4× H 8 × W 8 . Then, Gaussian noise is added to Z c to obtain the added noisy latent z c at timestep t. At the i-th layer of the denoising UNet, the hybrid attention output Z (i) hs is defined are as follow: Z (i) h = Softmax QK ⊤ √ d V + β Softmax Q K (i) o ⊤ √ d ! V (i) o , (5) where β is a hyperparameter controlling the contribution of the conditioned branch. Specifically, the query, key, and value matrices of the latent self-attention branch are defined as Q = z c W q , K = z c W k and V = z c W v while the condition branch is constructed from the masked model UNet features M ′(i) o where the corresponding key and value are given by K (i) o = M ′(i) o ˆ W k and V (i) o = M ′(i) o ˆ W v . 8S. Liu et al. InputTarget BridgeDiff (Ours) MGT Any2AnyTryonInputTarget MGT Any2AnyTryon BridgeDiff (Ours) Fig. 4: Qualitative comparisons on the DressCode dataset. Red circles highlight dif- ferences 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. The projection matrices ˆ W k and ˆ W v are initialized from W k and W v , respectively, and remain learnable during training. Flat Structure Constraint Module. For flat garment generation, the canon- ical flat layout cannot be reliably preserved when conditioning relies solely on textual descriptions. To address this issue, we focus on decoupling semantic ap- pearance control from structural layout constraints. We introduce the FSCM into the denoising UNet at specific stages to explicitly enforce flat garment lay- outs during generation. The module consists of a single flat-constraint atten- tion (FC-Attention) layer and a learnable linear projection layer. Specifically, we extract garment clues representation from the GCBM, which are denoted as F c ∈ R N×1024 . In parallel, the corresponding flat garment textual description 6 is processed by a text encoder, producing text embeddings T flat ∈ R 77×768 . To align with the dimensionality of the text embeddings, F c is further projected through a linear projection layer, resulting visual garment cues F ′ c ∈ R N×768 . The FC-Attention layer is designed to fuse visual garment cues with flat garment textual descriptions. For a single attention head m, the FC-Attention output is defined as: z (m) = Softmax Q (m) K (m) ⊤ √ d ! V (m) ,(6) The final FC-Attention output is obtained by concatenating the outputs of all G heads and applying a learnable linear projection W o : Z = Concat z (1) ,z (2) ,...,z (G) W o ,(7) 6 Inspired by the captioning convention used in CLIP pre-training [25], we employ a fixed textual template, “A flat-lay <category>,” to describe all target flat garment images. BridgeDiff for Virtual Try-Off9 Table 2: Quantitative results on the VITON-HD dataset over the full test set and category-specific subsets. ↑ indicates higher is better, ↓ indicates lower is better. Bold denotes the best result and underline denotes the second-best result. “–” indicates the metric is not reported. † denotes cross-dataset evaluation where the model is trained on DressCode and evaluated on VITON-HD. MethodFID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ TryOffDiff [36]22.548.5811.7771.8342.2724.87 TryOffAnyone [40]12.592.7712.5171.9034.4722.68 Any2AnyTryOn [9] 12.151.96 12.8674.2935.7222.80 Voost [16]10.06 2.48---- UniFit [46]12.58--77.5028.1020.20 BridgeDiff (Ours)9.081.5315.0077.4224.3818.69 MGT † [37]24.00 10.0610.6370.7747.4529.98 BridgeDiff (Ours) † 17.655.7512.8874.2435.0524.79 where Q is projected from the visual garment cues F ′ c , and K and V are projected from the flat garment textual embeddings T flat . The resulting FC-Attention out- put is denoted as Z fc ∈ R N×768 , which subsequently injected into the denoising UNet at selected stage. Specifically, at the j-th cross-attention layer, Z fc is in- corporated as an additional conditioning signal alongside the appearance-based textual description. Given a garment appearance prompt encoded by a text en- coder as T c and the corresponding hidden feature at this layer denoted as z c , The modified cross-attention output z j is defined as: z j = Softmax QK ⊤ √ d V + γ Softmax QK ⊤ fc √ d ! V fc , (8) where γ balances the contribution of structure guidance. The query is Q = z c W q . For the appearance-conditioned branch, K = T c W k and V = T c W v , while for the branch conditioned on the flat structure, K fc = Z fc W f k and V fc = Z fc W f v . The projection matrices W f k and W f v are initialized from W k and W v and re- main learnable. All other cross-attention layers in the denoising UNet use only textual appearance features T c , as in standard Text-to-Image (T2I). This injec- tion strategy decouples semantic appearance control from structural layout con- straints, preserving the native T2I generation capability of the denoising UNet. The explicit structural constraints of the FSCM strengthen flat garment struc- tures during the diffusion process, effectively stabilizing the structures without sacrificing appearance fidelity. 4 Experiment 4.1 Datasets and Metrics Our experiments are conducted on two public datasets: VITON-HD [3] and DressCode [21]. VITON-HD contains 13,679 high-resolution (1024× 768) upper- 10S. Liu et al. Input TryOffAnyoneTryOffDiff MGT Any2AnyTryon Target BridgeDiff (Ours) Fig. 5: Qualitative comparisons on the VITON-HD dataset. Red circles highlight dif- ferences 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. body garment image pairs, with 11,647 pairs for training and 2,032 pairs for testing. DressCode includes 53,792 high-resolution (1024× 768) full-body per- son–garment pairs, with 48,392 pairs for training and 5,400 pairs for testing, covering upper-body, lower-body, and dress categories. To evaluate the recon- struction quality in paired try-off settings, we adopt several widely used full- reference and perceptual metrics: SSIM [39], LPIPS [45], FID [23], KID [2], DISTS [6], and PSNR. 4.2 Implementation details Our framework is trained in a two-stage manner. In the first stage, the model follows MetaFormer [44] architecture. Following the design of ConvNeXt [18], each Feed-Forward Network (FFN) has an intermediate layer 4 times the hid- den dimension. We employ OpenCLIP ViT-H/14 [12] as the image encoder. In the second stage, both models are built upon the standard Stable Diffusion v1.5 [27]. For the model UNet, we employ SigLIP2 L/16-512 [35] and OpenAI ViT-L/14 [25] as the image encoder and text encoder, respectively. For the de- noising UNet, OpenAI ViT-L/14 is used as the text encoder. For second stage, we apply the DREAM [47] training strategy with the parameter set to p = 1. All experiments are conducted on 4 NVIDIA RTX 4090 GPUs. For each stage, we train two separate model versions on the VITON-HD and DressCode datasets independently. We use the AdamW optimizer [19] with a constant learning rate BridgeDiff for Virtual Try-Off11 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 compar- ison among different FSCM variants (M1–M3). MethodFID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ Effectiveness of the GCBM w/o GCBM9.37 1.51 15.0576.9124.9319.34 w/ GCBM (Ours)9.081.5315.0077.4224.3818.69 Effectiveness of the FSCM M1 (w/o T c )9.261.8915.1076.9324.1418.71 M2 (w/o F ′ c )9.461.6215.2277.1124.1118.96 M3 (w/o FSCM)9.421.6014.9077.0324.9619.19 w/ FSCM (Ours)9.081.5315.0077.4224.3818.69 of 1× 10 −5 and a weight decay of 0.01. The batch size is set per GPU, with 128 for the first stage and 8 for the second stage. Both stages are trained at a resolution of 512× 512, with 65k and 54k training steps, respectively. 4.3 Main Results We compare our method with recent VTOFF approaches. VTOFF-specific meth- ods are TryOffDiff [36], MGT [37], and TryOffAnyone [40]. Unified multi-task frameworks are Voost [16], Any2AnyTryOn [9], and UniFit [46]. Due to unavail- able implementations, results for Voost and UniFit are taken from their original papers. TryOffDiff and TryOffAnyone are evaluated only on VITON-HD, while MGT is tested in a cross-dataset setting. Under the same evaluation protocol, we also report comparable results on VITON-HD. For qualitative comparison, all methods follow their official inference instructions, and quantitative metrics are computed uniformly from the generated images. Results on the DressCode Dataset. Table 1 summarizes the quantitative results on the DressCode dataset over the full test set and category-specific sub- sets. Overall, BridgeDiff outperforms existing methods on the full test set and shows clear advantages on several garment categories, particularly on lower-body garments and dresses, which are more prone to side-view occlusions and struc- tural ambiguity. In such cases, BridgeDiff better preserves structural stability while maintaining visually continuous boundary details. For upper-body gar- ments, although BridgeDiff does not achieve the best scores on all perceptual metrics, it remains competitive when perceptual quality and structural consis- tency are jointly considered. Figure 4 presents qualitative comparisons on the DressCode dataset. Existing methods often suffer from structural distortions or visual discontinuities in occluded regions, whereas BridgeDiff produces flatter garment layouts with improved visual continuity. These quantitative and quali- tative improvements stem from the joint modeling of garment cues representation 12S. Liu et al. Model w/o GCBM Target Model M1 (w/o ) Target w/ FSCM (Ours) M2 (w/o ) M3 (w/o FSCM) w/ GCBM (Ours) Fig. 6: Qualitative comparisons on the VITON-HD dataset. Left: GCBM qualita- tive analysis. Right: FSCM qualitative analysis. Red circles highlight differences in local regions across different methods. Zooming in provides a clearer view of these differences. and explicit flat structure constraints, which stabilize garment structure during diffusion and enable continuous detail inference in occluded areas. Results on the VITON-HD Dataset. Table 2 summarizes the quantita- tive results on the VITON-HD dataset. Overall, BridgeDiff achieves the best performance across most evaluation metrics, validating its overall advantages in perceptual quality and structural stability. Under cross-dataset evaluation, BridgeDiff still maintains a notable performance margin, highlighting its robust generalization to unseen data distributions. Figure 5 presents qualitative com- parisons on the VITON-HD dataset. Existing methods often struggle to main- tain geometric stability of garments and visual continuity in occluded regions, which frequently leads to structural distortions and misaligned boundaries. In contrast, BridgeDiff is able to generate structurally stable garment results, while exhibiting superior visual continuity in occluded areas. Both quantitative and qualitative results demonstrate that incorporating explicit garment cues repre- sentation together with structured flat garment constraints facilitates the gener- ation of garment results that are more stable and continuous in both structure and visual appearance. 4.4 Ablation Studies and Analysis To analyze the effectiveness of the core components in BridgeDiff, we conduct two groups of controlled ablation studies on the VITON-HD dataset. All experiments follow the same settings as the main results unless otherwise specified. Due to space limitations, additional ablation studies and detailed analyses are provided in the Appendix. Effectiveness of the GCBM. To verify the effectiveness of GCBM, we remove the explicit garment cues and replace them with the textual conditions from BridgeDiff for Virtual Try-Off13 Model w/ FSCM (Ours) w/o FSCM Model w/o GCBM w/ GCBM (Ours) Fig. 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 high- light differences in local regions across different methods. the original FSCM, while also eliminating the class embedding. The ablation variant is denoted as w/o GCBM, while the model equipped with explicit garment cues is denoted as w/ GCBM. As shown in the upper part of Table 3, introducing GCBM leads to consistent improvements on most perceptual and structure-related metrics, indicating that the model is able to generate more continuous visual features under limited conditioning. As illustrated in the left part of Figure 6, without GCBM, the model often produces garments with locally plausible textures but exhibits noticeable visual discontinuities in occluded or weakly observed regions. In contrast, by introducing garment cues, the model is able to naturally infer unseen regions and generate visually continuous garment features. This qualitative comparison highlights the role of GCBM in generating garment cues, which effectively improves the continuity of appearance modeling under limited or partial observations. Effectiveness of the FSCM. To evaluate the effectiveness of FSCM, we design three ablation variants while keeping all other components unchanged. Specifi- cally, M1 removes T c and conditions the model only on the image-encoder-based garment features F ′ c , while M2 removes F ′ c and retains only the flat garment textual description T c as the structural condition. These two variants evaluate whether a single garment condition is sufficient to stabilize garment structure. In M3, the FSCM module is entirely removed and replaced by a prompt-level aug- mentation, where additional flat garment descriptions T ′ c are prepended to the original garment appearance prompt 7 . This setting assesses whether such naive prompt-level augmentation can maintain garment structural stability while pre- serving the correctness and continuity of visual information. As shown in the lower part of Table 3, the complete FSCM demonstrates the best overall per- formance, leading across multiple structure-related metrics, which indicates its superior effectiveness in maintaining the global structural stability of flat gar- ments. Although its PSNR and LPIPS are slightly lower than those of some variants, this is expected, as the flat-structure constraint module emphasizes 7 The added textual description mainly specifies key flat garment appearance at- tributes. Concretely, we prepend “A flat-lay <category>,” to the original garment appearance prompt for all target flat garment images. 14S. Liu et al. (a) User study results of VITON-HD (b) User study results of DressCode Fig. 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. canonical flat garment geometry rather than pixel-level similarity or local per- ceptual consistency. At the same time, its advantages in distribution-level metrics such as FID further indicate that the use of the complete FSCM enables more stable preservation of the overall garment structure compared to other variants. As shown in the right part of Fig. 6, relying on a single garment condition of- ten leads to local wrinkles, structural errors, or discontinuous visual features. Even M3, which uses prompt-level augmentation, cannot consistently maintain structural stability and still produces subtle wrinkles, geometric inconsistencies, or incorrect fine-grained details. In contrast, FSCM leverages complementary visual cues and explicit textual constraints to ensure structural stability in flat garments and generate continuous visual results. Attention Heatmap Visualization. To analyze the effects of the proposed modules, we visualize the attention weights in the denoising UNet at the 50th denoising step. Specifically, we extract the attention distribution from the down blocks.0.attentions.1 layer, where each spatial position of the denoising latent attends to the target garment features. The aggregated attention weights are mapped into spatial heatmaps for visualization. As shown in Fig. 7, the full model with FSCM assigns higher attention weights around garment structural boundaries, indicating improved modeling of flat garment layouts. In contrast, the variant without FSCM mainly focuses on locally visible regions. For GCBM, the model maintains strong attention responses in occluded areas, while remov- ing GCBM leads to noticeably weaker attention in these regions. 4.5 User Study To evaluate the perceptual quality and structural consistency of generated gar- ments, we conduct a questionnaire-based user study. Participants are shown anonymized and shuffled results from different methods under the same input model image, with multiple random seeds for fairness. Each result is evaluated based on Visual Realism, Garment Structure Consistency, Occluded Region Continuity, and Overall Preference. The first three criteria are rated on a three-point scale (0–2), while overall preference is collected as a single-choice selection. Results are aggregated by the total number of evaluated samples. As shown in Fig. 8, our method achieves consistently higher ratings on BridgeDiff for Virtual Try-Off15 both VITON-HD and DressCode, with more participants selecting our results as their preferred outputs. 5 Conclusion and Future Work We proposed BridgeDiff, which explicitly bridges the gap between human-centric garment observations and canonical flat garment synthesis. By introducing gar- ment cue representations and injecting explicit flat-structure constraints during denoising, BridgeDiff effectively alleviates visual discontinuities in unobservable regions and improves structural stability compared to text-only constrained ap- proaches. Extensive experimental results demonstrate that BridgeDiff achieves competitive performance across multiple garment categories, achieving better visual continuity in occluded regions and enhanced overall structural stability. Nevertheless, the method can still be imprecise or incomplete under severe oc- clusions or extreme poses. In future work, we plan to train on larger datasets to further improve the model’s ability to infer visually continuous garment regions. References 1. Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 2. Bińkowski, M., Sutherland, D.J., Arbel, M., Gretton, A.: Demystifying mmd gans. arXiv preprint arXiv:1801.01401 (2018) 3. Choi, S., Park, S., Lee, M., Choo, J.: Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 14131–14140 (2021) 4. Choi, Y., Kwak, S., Lee, K., Choi, H., Shin, J.: Improving diffusion models for authentic virtual try-on in the wild. In: European Conference on Computer Vision. p. 206–235. Springer (2024) 5. Chong, Z., Dong, X., Li, H., Zhang, S., Zhang, W., Zhang, X., Zhao, H., Jiang, D., Liang, X.: Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886 (2024) 6. Ding, K., Ma, K., Wang, S., Simoncelli, E.P.: Image quality assessment: Unify- ing structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence 44(5), 2567–2581 (2020) 7. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020) 8. Gou, J., Sun, S., Zhang, J., Si, J., Qian, C., Zhang, L.: Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 7599–7607 (2023) 9. Guo, H., Zeng, B., Song, Y., Zhang, W., Liu, J., Zhang, C.: Any2anytryon: Leverag- ing adaptive position embeddings for versatile virtual clothing tasks. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision. p. 19085– 19096 (2025) 16S. Liu et al. 10. Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems 33, 6840–6851 (2020) 11. Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022) 12. Ilharco, G., Wortsman, M., Carlini, N., Taori, R., Dave, A., Shankar, V., Namkoong, H., Miller, J., Hajishirzi, H., Farhadi, A., et al.: Openclip. Zenodo (2021) 13. Kim, J., Gu, G., Park, M., Park, S., Choo, J.: Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 8176– 8185 (2024) 14. Kingma, D.P., Welling, M.: Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013) 15. Lee, S., Gu, G., Park, S., Choi, S., Choo, J.: High-resolution virtual try-on with misalignment and occlusion-handled conditions. In: European Conference on Com- puter Vision. p. 204–219. Springer (2022) 16. Lee, S., Kwak, J.g.: Voost: A unified and scalable diffusion transformer for bidi- rectional virtual try-on and try-off. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers. p. 1–11 (2025) 17. Liu, J., He, Z., Wang, G., Li, G., Lin, L.: One model for all: Unified try-on and try-off in any pose via llm-inspired bidirectional tweedie diffusion. arXiv preprint arXiv:2508.04559 (2025) 18. Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 11976–11986 (2022) 19. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017) 20. Morelli, D., Baldrati, A., Cartella, G., Cornia, M., Bertini, M., Cucchiara, R.: Ladi- vton: Latent diffusion textual-inversion enhanced virtual try-on. In: Proceedings of the 31st ACM international conference on multimedia. p. 8580–8589 (2023) 21. Morelli, D., Fincato, M., Cornia, M., Landi, F., Cesari, F., Cucchiara, R.: Dress code: High-resolution multi-category virtual try-on. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 2231–2235 (2022) 22. Mou, C., Wu, Y., Wu, W., Guo, Z., Zhang, P., Cheng, Y., Luo, Y., Ding, F., Zhang, S., Li, X., et al.: Dreamo: A unified framework for image customization. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers. p. 1–12 (2025) 23. Parmar, G., Zhang, R., Zhu, J.Y.: On aliased resizing and surprising subtleties in gan evaluation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 11410–11420 (2022) 24. Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 4195–4205 (2023) 25. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. p. 8748–8763. PmLR (2021) 26. Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text- conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022) BridgeDiff for Virtual Try-Off17 27. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 10684–10695 (2022) 28. Shen, F., Jiang, X., He, X., Ye, H., Wang, C., Du, X., Li, Z., Tang, J.: Imagdressing- v1: Customizable virtual dressing. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, p. 6795–6804 (2025) 29. Shen, F., Tang, J.: Imagpose: A unified conditional framework for pose-guided person generation. Advances in neural information processing systems 37, 6246– 6266 (2024) 30. Shen, F., Ye, H., Zhang, J., Wang, C., Han, X., Wei, Y.: Advancing pose-guided image synthesis with progressive conditional diffusion models. In: The Twelfth In- ternational Conference on Learning Representations (2024), https://openreview. net/forum?id=rHzapPnCgT 31. Shen, F., Yu, J., Wang, C., Jiang, X., Du, X., Tang, J.: Imaggarment-1: Fine-grained garment generation for controllable fashion design. arXiv preprint arXiv:2504.13176 (2025) 32. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics. In: International conference on machine learning. p. 2256–2265. pmlr (2015) 33. Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020) 34. Team, G., Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., Perrin, S., Matejovicova, T., Ramé, A., Rivière, M., et al.: Gemma 3 technical report. arXiv preprint arXiv:2503.19786 (2025) 35. Tschannen, M., Gritsenko, A., Wang, X., Naeem, M.F., Alabdulmohsin, I., Parthasarathy, N., Evans, T., Beyer, L., Xia, Y., Mustafa, B., et al.: Siglip 2: Multilingual vision-language encoders with improved semantic understanding, lo- calization, and dense features. arXiv preprint arXiv:2502.14786 (2025) 36. Velioglu, R., Bevandic, P., Chan, R., Hammer, B.: Tryoffdiff: Virtual-try-off via high-fidelity garment reconstruction using diffusion models. arXiv preprint arXiv:2411.18350 (2024) 37. Velioglu, R., Bevandic, P., Chan, R., Hammer, B.: Mgt: Extending virtual try- off to multi-garment scenarios. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 6039–6048 (2025) 38. Wang, B., Zheng, H., Liang, X., Chen, Y., Lin, L., Yang, M.: Toward characteristic- preserving image-based virtual try-on network. In: Proceedings of the European conference on computer vision (ECCV). p. 589–604 (2018) 39. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004) 40. Xarchakos, I., Koukopoulos, T.: Tryoffanyone: Tiled cloth generation from a dressed person. arXiv preprint arXiv:2412.08573 (2024) 41. Xu, Y., Gu, T., Chen, W., Chen, A.: Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, p. 8996–9004 (2025) 42. Yang, Z., Li, Y., He, S., Li, X., Xu, Y., Dong, J., Du, Y.: Omnivton: Training-free universal virtual try-on. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. p. 16702–16711 (2025) 43. Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compati- ble image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023) 18S. Liu et al. 44. Yu, W., Luo, M., Zhou, P., Si, C., Zhou, Y., Wang, X., Feng, J., Yan, S.: Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 10819–10829 (2022) 45. Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE conference on computer vision and pattern recognition. p. 586–595 (2018) 46. Zhang, W., Jin, Y., Li, X., Zhang, Y., Cong, X., Wang, C., Qiao, F., et al.: Unifit: Towards universal virtual try-on with mllm-guided semantic alignment. arXiv preprint arXiv:2511.15831 (2025) 47. Zhou, J., Ding, T., Chen, T., Jiang, J., Zharkov, I., Zhu, Z., Liang, L.: Dream: Diffusion rectification and estimation-adaptive models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 8342– 8351 (2024) 48. Zhu, L., Yang, D., Zhu, T., Reda, F., Chan, W., Saharia, C., Norouzi, M., Kemelmacher-Shlizerman, I.: Tryondiffusion: A tale of two unets. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. p. 4606–4615 (2023) BridgeDiff for Virtual Try-Off19 Appendix Supplementary Material This supplementary material provides an extended and in-depth exploration of the experiments and methodologies presented in the main paper. Section A in- troduces additional background and defines key symbols to facilitate a clearer understanding of the proposed framework. Section B describes the training strat- egy and inference pipeline in detail, covering important implementation choices and practical considerations. Section C presents extended experimental analy- ses, including more comprehensive ablation studies. Section D provides a detailed description of the user study procedure. Section E explains the procedure used to generate garment-related textual descriptions. Section F showcases a broader range of qualitative comparison examples with state-of-the-art methods, offering further insights into the strengths of our approach. A Preliminaries and Notations A.1 Preliminaries Latent Diffusion Models. The core idea of Latent Diffusion Models (LDMs) [27] reduce the computational cost of diffusion-based generative models by con- ducting the diffusion process in a compact latent space. Given an input image x 0 , a pre-trained Variational Autoencoder (VAE) [14] is first used to encode the image into a latent representation, enabling efficient training and inference while maintaining high generation quality. An LDM mainly consists of a denois- ing UNet ε θ (·,t) and a VAE composed of an encoderE and a decoderD. During training, Gaussian noise is progressively added to the clean data according to the forward diffusion process, producing a noisy sample at timestep t: z t = √ ̄α t z 0 + √ 1− ̄α t ε,(9) where z 0 = E (x 0 ) denotes the latent representation of the input image x 0 ob- tained from the pre-trained VAE encoder E, ε ∼ N (0, 1) is standard Gaussian noise, and ̄α t is a predefined function of t that determines the diffusion schedule. The denoising UNet is trained to predict the injected noise from the noisy input z t , and the training objective is defined as: L LDM := E z 0 ,ε,t h ∥ε− ε θ (z t ,t)∥ 2 2 i (10) where t∈1,...,T denotes the timestep of the diffusion process. At inference time, samples drawn from the latent distribution are iteratively denoised and finally mapped back to the image space through a single forward pass of the VAE decoder D. 20S. Liu et al. Table 4: Notation Definition SymbolDefinition 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) εGaussian noise t Timestep Z m Model image encoded via VAE I m Encoded feature of model image (stage 2) I ′ m Projected I m T m Encoded textual description of model M c Cloth-agnostic mask M o Intermediate features of model UNet M ′ o Features injected into denoising UNet Z c Target garment encoded via VAE z c Z c with added noise T flat Encoded textual description of flat garment F ′ c Projected output of garment condition bridge module (GCBM) GNumber of attention heads in flat structure constraint module (FSCM) Z fc Final output of flat structure constraint module (FSCM) T c Encoded appearance description of garment ˆz c Rectified input of target garment ˆεRectified new target noise λBalancing hyperparameter for model UNet hybrid cross-attention βBalancing hyperparameter for denoising UNet hybrid attention γBalancing hyperparameter for denoising UNet hybrid cross-attention wClassifier-Free Guidance (CFG) weight Diffusion Rectification and Estimation-Adaptive Models. Diffusion Rec- tification and Estimation-Adaptive Models (DREAM) [47] is a training strategy designed to alleviate the trade-off between perceptual quality and pixel-level distortion in conditional generation tasks. During training, the diffusion model computes a rectified input and a corresponding target as follows: ˆz t = z t + √ 1− ̄α t λ t ∆ε t,θ ,ˆε = ε + λ t ∆ε t,θ ,(11) where ∆ε t,θ = ε− StopGradient(ε θ (z t ,t)), λ t = ( √ 1− ̄α t ) p with p being a hyper- parameter, ˆz t denotes the rectified input and ˆε is the new target. The training objective is defined as: L DREAM := E z t ,ˆε,t h ˆε− ε θ (ˆz t ,t) 2 2 i , (12) During inference, the standard reverse diffusion process [33] is still adopted. DREAM improves training accuracy and convergence efficiency, although each BridgeDiff for Virtual Try-Off21 training step requires an additional forward pass, slightly increasing the compu- tational cost. A.2 Notations All symbols and their corresponding definitions are provided in detail in Table 4. B Training and Inference Details The proposed Flat Structure Constraint for Conditional Diffusion is optimized using only a single training objective, without introducing any auxiliary losses. Specifically, for the model UNet, all parameters as well as the image projection layers are fully trainable. In contrast, for the denoising UNet, we only update the class embedding layer, all linear projection layers within the FSCM, the key and value projection matrices W f k and W f v at the specific j-th cross-attention layer, and the projection matrices ˆ W k and ˆ W v in all hybrid attention layers. All remaining parameters are kept frozen during training. The training objective follows the standard diffusion noise learning formu- lation. Given the model features I ′ m , model text descriptions T m , clean model latent representations Z m , garment appearance text embeddings T c , garment cues representation F c , projected garment cues representation F ′ c , flat garment text embeddings T flat , Gaussian noise ε ∼ N (0, 1), added noisy latent z c and diffusion timestep t, the diffusion loss is defined as: L diff = E I ′ m ,T m ,Z m ,T c ,F c ,F ′ c ,T flat ,ε,z c ,t h ˆε− ε θ ( ˆz c , I ′ m ,T m ,Z m ,T c ,F c ,F ′ c ,T flat ,t) 2 2 i , (13) where ε θ (·) denotes the noise prediction network, ˆz c denoted the rectified input and ˆε is the new target [47]. During inference, we adopt Classifier-Free Guidance (CFG) [11] to perform conditional sampling. Specifically, given the noisy latent ε, the conditional noise prediction ε θ (ε,I ′ m ,T m ,Z m ,T c ,F c ,F ′ c ,T flat ,t) and the unconditional noise pre- diction ε θ (ε,t), the final guided noise estimate is obtained as: ˆε o = ε θ (ε,t) + w ε θ (ε,I ′ m ,T m ,Z m ,T c ,F c ,F ′ c ,T flat ,t)− ε θ (ε,t) (14) where w, often referred to as the guidance scale is a scalar that controls the strength of conditioning. C Additional Experimental Results and Analysis To analyze the effectiveness of the core components in BridgeDiff, we conduct multiple controlled ablation and comparative studies on the VITON-HD dataset. All experiments follow the same settings as the main paper unless otherwise 22S. Liu et al. Table 5: Ablation study on FSCM insertion positions on the VITON-HD dataset. ↑ indicates higher is better, ↓ indicates lower is better. Bold denotes the best result. MethodFID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ Down 19.421.8615.0077.4924.1818.82 Down 29.321.8815.0377.2424.2718.77 Mid9.251.8415.0777.3624.2218.77 Up 19.421.8115.0377.4824.2018.73 Up 29.031.6715.0577.0024.4218.71 Up 39.351.8514.9277.0624.3718.71 All9.511.9114.8477.0924.5318.81 Down 0 (Ours)9.081.5315.0077.4224.3818.69 noted. Specifically, we investigate the impact of the FSCM insertion position on generation quality (see Section C.1), compare BridgeDiff with UniFit under the virtual try-off (VTOFF) setting to analyze how different methods address the challenge of focusing on garment regions (see Section C.2), and analyze the effect of removing the model textual description T m and garment appearance textual description T c (see Section C.3). Finally, we investigate the sensitivity of key hyperparameters of the proposed framework (see Section C.4). Taken together, these analyses systematically demonstrate the rationality and effectiveness of the key design choices in BridgeDiff. C.1 FSCM Insertion Position Analysis This subsection investigates how the insertion point of the FSCM affects the stability and visual quality of generated flat garments. In our experiments, we leverage the layer naming convention from unet.attn_processors to denote the FSCM insertion positions. Specifically, Down 0, Down 1, Down 2 correspond to the three consecutive downsampling stages of the UNet, while Up 1, Up 2, Up 3 indicate the respective upsampling stages. The Mid stage refers to the intermediate bottleneck layer connecting the encoder and decoder. In addition, the variant All refers to inserting FSCM into all available stages simultaneously. This setup allows us to systematically analyze the effect of structural guidance on global garment geometry and visual continuity, providing insights into where explicit flat-structure constraints are most effective. As shown in Table 5, Down 0 achieves the most balanced performance across distribution-level, perceptual, and low-level reconstruction metrics. Specifically, Down 0 yields the lowest KID and DISTS, while maintaining competitive FID and SSIM, indicating improved global structural stability and visual continu- ity. In comparison, inserting FSCM at a slightly later downsampling stage (e.g., Down 1) achieves relatively strong performance on SSIM and LPIPS, but does not provide clear advantages on distribution-level metrics (e.g., FID and KID), suggesting that structural constraints introduced at this stage are less effective in stabilizing global garment structure. Furthermore, inserting FSCM at later BridgeDiff for Virtual Try-Off23 InputUp 1MidDown 2Down 1Up 2TargetDown 0 (Ours) All Up 3 Fig. 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. stages (e.g., Up 2) slightly improves FID and PSNR, but leads to inferior per- formance on perceptual or structure-related metrics, indicating less consistent garment structure. Similarly, inserting FSCM into all layers results in degraded performance across most metrics, implying that overly strong structural con- straints may hinder effective modeling during the diffusion process. Overall, these results demonstrate that introducing flat structure constraints at the earliest stages of the diffusion process is more effective for stabilizing flat-garment ge- ometry while preserving overall visual quality. Qualitative comparisons in Fig. 9 further highlight the impact of FSCM insertion positions. In the first three ex- amples, alternative insertion strategies frequently introduce unnatural wrinkles, fragmented textures, or visually discontinuous regions, especially around areas with complex folds, hair occlusions, or garment boundaries. Even when the over- all layout appears plausible, these variants often fail to maintain continuous and coherent visual information under partial observations. More interestingly, the fourth example reveals a subtle yet informative case. Although only a small and easily overlooked garment region is visible near the model’s lower-left boundary (i.e., the upper garment region highlighted by the red circle), most insertion variants are able to capture this weak visual cue and generate different hem configurations. However, only our early insertion strategy produces a result that is closest to the ground truth, correctly inferring the garment structure while maintaining visual continuity. This observation further demonstrates that GCBM can effectively generate garment cue representations and interact with the flat textual representation at an early stage. However, incorporating criti- cal structural information before geometric ambiguities accumulate during the 24S. Liu et al. diffusion process is particularly important for maintaining consistent garment structure. C.2 Comparison with UniFit’s Spatial Attention Focusing Loss In VTOFF tasks, diffusion models often attend to regions in the model image that are irrelevant to the target garment (e.g.,background or non-garment body parts), which degrades fine-grained garment modeling and introduces visual arti- facts. To address this issue, UniFit [46] is inspired by DreamO [22] and introduces a spatial attention focusing loss that explicitly regularizes cross-attention maps to encourage attention concentration on garment regions. Specifically, UniFit ag- gregates cross-attention responses between denoising queries and reference image (i.e., model image) keys across attention heads and reference tokens. Finally, a spatial attention map M is aligned with a pre-extracted cloth-agnostic mask M c from the model image via an MSE loss: L focus = 1 N R N L N L X j=1 N R X i=1 M j i − M c,i 2 2 .(15) where M j i denotes the response map for the i-th model image in the j-th at- tention layer, M c,i is the cloth-agnostic mask corresponding to the i-th model image, and N R and N L represent the number of model images and attention layers, respectively. In contrast, we adopt a simpler and more efficient strategy. Instead of introducing an additional optimization objective, we directly leverage the cloth-agnostic mask to extract clean garment regions from the model image before attention computation. This design avoids interference from irrelevant regions without extra training losses, leading to a simpler and more direct opti- mization process. In our experiments, we compare our method with UniFit and denote the variant with the attention focusing loss as w/ Loss, and the variant that directly extracts garment regions as w/o Loss. As shown in the upper part of Table 6, directly extracting garment re- gions consistently outperforms the UniFit variant using the spatial attention focusing loss across nearly all metrics. This indicates that directly extracting garment regions via the cloth-agnostic mask is more effective than relying on at- tention regularization for suppressing irrelevant information, leading to cleaner garment reconstruction and improved perceptual and structural fidelity. Quali- tative comparisons in Figure 10a further illustrate this effect. When the model wears semi-transparent garments, BridgeDiff is able to accurately generate the clean underlying garment, whereas UniFit tends to retain colors from the visible body. Additionally, some visual information around the garment is also incor- porated into the generated clothing. We believe that since the cross-attention uses softmax, non-garment regions still receive non-zero attention scores, lead- ing to extra details from irrelevant areas at the garment. In contrast, directly masking out irrelevant regions ensures the generated garments focus solely on the target cloth, improving both visual clarity and structural consistency. It is BridgeDiff for Virtual Try-Off25 Table 6: Ablation study results on the VITON-HD dataset. ↑ indicates higher is better, ↓ 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. MethodFID↓ KID↓ PSNR↑ SSIM↑ LPIPS↓ DISTS↓ Comparison with UniFit’s Spatial Attention Focusing Loss w/ Loss (UniFit) 9.781.9914.9277.3724.6718.66 w/o Loss (Ours)9.081.5315.0077.4224.3818.69 Impact of Textual Descriptions w/o Text9.061.5715.0077.0724.4218.76 w/ Text (Ours)9.081.5315.0077.4224.3818.69 worth noting that our direct garment region extraction is extremely simple and effective, and more importantly, it does not introduce any additional training objectives. C.3 Impact of Textual Descriptions To study the role of textual descriptions in our model, we analyze the impact of removing the model textual description T m and the garment appearance textual description T c on performance. Specifically, for the ablation study, we define the variant without textual inputs as w/o Text. In this variant, the model UNet receives the encoded model image, which is first processed by the image encoder and then projected through a linear layer before being injected into the UNet. This injection procedure follows the same design as the IP-Adapter [43]. Mean- while, the denoising UNet no longer receives the garment appearance textual description T c . Instead, only the FSCM is injected into the cross-attention lay- ers, while all other cross-attention inputs are replaced with empty text tokens. In contrast, the variant using both the model textual description T m and the garment appearance textual description T c is defined as w/ Text. As shown in the lower part of Table 6, removing both the model textual description T m and the garment appearance description T c does not lead to a significant degradation in overall quantitative performance. This surprising result suggests that the core structural and visual information required for gar- ment synthesis are largely captured by the FSCM and the garment cues, rather than relying heavily on textual guidance. Nevertheless, the noticeable drop in SSIM indicates that textual descriptions still contribute to improving structural consistency in specific regions (e.g., garment necklines or waists). Qualitative comparisons in Figure 10b further illustrate the role of textual descriptions. Without text guidance, the model may introduce visually plausible but semanti- cally irrelevant details in ambiguous garment regions, especially near boundaries areas. In contrast, incorporating textual descriptions encourages the model to 26S. Liu et al. Model w/ Loss (UniFit) Target w/o Loss (Ours) (a) Qualitative comparison of UniFit attention loss vs. our direct garment region extraction. Model w/o Text Target w/ Text (Ours) (b) Qualitative comparison illustrating the im- pact of textual descriptions. Fig. 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. focus on the currently visible garment semantics, leading to more coherent and semantically consistent synthesis results. C.4 Hyperparameter Sensitivity We validate two key hyperparameters in our method. Specifically, we investi- gate the effects of p, the hyperparameter in DREAM [47], which is designed to alleviate the trade-off between perceptual quality and pixel-level distortion in conditional generation tasks, and w, the CFG [11] weight that balances condi- tional faithfulness and sample diversity during generation. Diffusion Rectification and Estimation-Adaptive Models. As described in Sec. A.1, we incorporate DREAM to improve the training of conditional dif- fusion models. The hyperparameter p controls the strength of DREAM. Specif- ically, DREAM is disabled and the training objective reduces to the standard diffusion formulation when p = ∞, whereas DREAM is activated when p takes finite values. We qualitatively analyze the effect of DREAM. As shown in the upper part of Figure 11, a small p leads to overly smooth outputs, while a large p introduces excessive high-frequency noise, resulting in noticeable wrinkles and local irregularities in the garments. This highlights the importance of properly selecting the strength of DREAM to balance perceptual fidelity and distortion regulation. Table 7 presents the results of training with different p values on the BridgeDiff for Virtual Try-Off27 Input= 2= Input = 6.0 = 1.5 = 1.0 (Ours)= 0.5= 0.0 = 5.0 = 4.0= 3.0 (Ours)= 2.0 = 1.0 Fig. 11: Qualitative comparisons on the VITON-HD dataset. Upper: DREAM qual- itative analysis. Lower: CFG qualitative analysis. Red circles highlight differ- ences in local regions across different methods. Zooming in provides a clearer view of these differences. VITON-HD dataset. As p increases, distribution-level metrics (such as FID and KID) initially improve but increase again at larger p values, while LPIPS exhibits a similar trend. Overall, p = 1.0 achieves a good balance across all metrics. Classifier-Free Guidance. CFG balances between conditional and uncondi- tional generation, guiding the model to produce results that better satisfy the conditioning constraints while maintaining higher visual fidelity. To evaluate the effect of CFG on generation quality, we conduct inference strengths of 1.0, 2.0, 3.0, 4.0, and 5.0, while keeping all other experimental settings fixed, where w = 1.0 corresponds to standard conditional generation. As shown in the lower part of Figure 11, increasing the CFG strength im- proves image detail and fidelity. However, when the strength exceeds 3.0, the generated images exhibit severe artifacts and high-frequency noise, resulting in degraded visual quality. In our experiments, we found that a CFG strength be- tween 2.0 and 3.0 produces the most realistic and natural results. Therefore, we set the CFG strength to 3.0 for all experiments. D User Study Example Questionnaire To further evaluate the perceptual quality and garment structural consistency of our generated results, we conduct a user study using a questionnaire-based approach. The study is designed to be fully anonymous and adheres to common academic research ethics, collecting no personal or sensitive information. All par- ticipants are volunteers, and all results are recorded and analyzed anonymously. In the questionnaire, participants are presented with a series of images showing 28S. Liu et al. Table 7: Ablation results of different p in DREAM on the VITON-HD dataset. p =∞ means DREAM is disabled. ↑ indicates higher is better, ↓ indicates lower is better. Bold denotes the best result. pFID↓KID↓PSNR↑SSIM↑LPIPS↓DISTS↓ 0.011.923.5814.9977.5825.0920.51 0.59.602.0414.9677.4124.6019.14 1.0 (Ours)9.081.5315.0077.4224.3818.69 1.59.131.5814.8976.8624.5118.66 2.09.421.8414.8576.6624.4418.75 ∞9.151.6314.8777.2824.6418.66 the output of different methods for the same input model image. The generated images are anonymized and shuffled to prevent bias, and multiple random seeds are used to ensure fair comparison across all state-of-the-art methods. Partici- pants are asked to rate each result based on four criteria: Visual Realism, Gar- ment Structure Consistency, Occluded Region Continuity, and Overall Preference. Ratings are provided on a simple three-point scale (0 – Poor, 1 – Acceptable, 2 – Good), and overall preference is collected as a single-choice selec- tion among the competing methods. Participants are instructed to evaluate the images subjectively, considering visual realism, plausibility of garment structure, and continuity in occluded regions. An example of the questionnaire interface for a single image set is shown in Figure 13, illustrating how participants are asked to rate each method’s output. E Textual Description Generation To obtain fine-grained and controllable garment descriptions from model im- ages, we adopt a multi-stage text description generation pipeline assisted by multiple Multimodal Large Language Models (MLLMs). Unlike approaches that rely on free-form natural language descriptions, we first aim to construct a visu- ally grounded and semantically consistent structured garment attribute space. The attribute discovery and selection pipeline is illustrated in the left part of Fig. 12. Specifically, we introduce several state-of-the-art MLLMs, including ChatGPT [1] and Gemini [34], to collaboratively discover candidate garment attributes. Given a model image, the models are prompted to analyze garment appearance elements that are visually discriminative and structurally observable, while explicitly excluding color, texture, and pattern information. By aligning and comparing the candidate attributes proposed by different MLLMs, we retain only those attributes that reach consensus across multiple models and can be stably associated with visual appearance. This process results in a compact yet expressive set of garment attributes, which serves as the foundation of the struc- tured garment description space. Rather than producing unconstrained text, Qwen3-VL is guided to generate attribute-consistent descriptions following the BridgeDiff for Virtual Try-Off29 Model Image You are a fashion design expert. Given the garment worn by the model, please summarize the key elements that can be used to describe the garment from the perspective of observable structural appearance. Color information and any texture or pattern details should be ignored. Prompt Cloth Type × √ ∩ Layering Form Shape Fit Type Waist Type Cloth Length Hem WidthWaist Position Neckline Type Sleeve Length √ √ √ √ √ √ Hem Position Hem Geometry Neck Depth Component Symmetry × × × × × × ∩ Intersect System: You are a meticulous and highly skilled fashion annotation expert specialized in garment reconstruction and flat-lay clothing representation. Your task is to analyze a garment and reconstruct it as an independent, flat-lay product. Ignore any body, pose, or styling information. Focus only on the garment itself. Text: The garment should be imagined as laid flat, symmetrically arranged, with natural garment structure preserved. Reconstruct the garment as a flat-lay clothing item based solely on visual attributes present in the image. The clothing should be described as an independent product, not worn by a model. Absolutely do NOT generate colors, background, textures, or patterns. Do not use newline characters. Please Generate strictly in the following format: Reconstruct a <Cloth Type> with hem <Hem Width>, neckline <Neckline Type>, waist <Waist Type>, sleeve <Sleeve Length>, cloth <Cloth Length>, fit <Fit Type>. Image: : Reconstruct a knit with hem straight, neckline round, waist regular, sleeve length three quarter, cloth length short, fit regular. : Reconstruct a knit with hem straight, neckline round, waist regular, sleeve length three quarter, cloth length short, fit regular. User: Fig. 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 control- lable garment description generation. predefined template, ensuring interpretability and controllability of the gener- ated results. The interaction between the structured prompt and Qwen3-VL is shown in the right part of Fig. 12 The final generated garment appearance description is denoted as T c 8 . To accommodate different usage scenarios, we further perform lightweight text transformations on T c . Specifically, by replacing the leading keyword “Re- construct” with “A model is wearing”, we obtain a model-centric garment de- scription T m . Subsequently, by removing the subject phrase, we derive a flat-lay oriented garment description T flat , which follows a unified format of “A flat-lay <category>” and retains only the Cloth Type as the garment category indica- tor. These transformations adapt the description to different semantic perspec- tives without altering the underlying attribute semantics. F Qualitative Comparison with SOTA Methods To further evaluate our method, we provide additional qualitative comparisons with state-of-the-art (SOTA) methods on the DressCode and VITON-HD datasets. As shown in Fig. 14, we present results for Dresses on the DressCode dataset; Fig. 15 shows results for Lower-Body (first three rows) and Upper-Body (last three rows) on DressCode; and Fig. 16 shows results on the VITON-HD dataset. 8 Here, we interpret T c as the garment appearance description prior to being processed by the text encoder. Although T c is defined in Table 4 as the appearance descrip- tion of the garment after text encoding, we adopt this simplified interpretation to avoid introducing additional variables. The same notation convention is subsequently applied to T m and T flat . 30S. Liu et al. User Study Sample - VITON-HD (Sample ID: 13556_00) ModelCloth Method AMethod BMethod CMethod DMethod E Q1: Visual Realism Description: How realistic does the result look? Hint: Consider whether the image looks like a real product photo, and whether there are obvious artifacts, unnatural textures, or visual distortions. Method:Method AMethod BMethod CMethod DMethod E Score:n 0n 1n 2n 0n 1n 2n 0n 1n 2n 0n 1n 2n 0n 1n 2 Q2: Garment Structure Consistency Description: Is the garment structure reasonable and consistent with common e-commerce clothing images? Hint: Pay attention to whether the garment shape is correct, whether it has a flat appearance, and whether there is obvious distortion or excessive wrinkling. Method:Method AMethod BMethod CMethod DMethod E Score:n 0n 1n 2n 0n 1n 2n 0n 1n 2n 0n 1n 2n 0n 1n 2 Q3: Occluded Region Continuity Description: For occluded or partially visible garment regions, does the result have visual continuity and reasonability? Hint: Focus on areas where the bottom of the garment is tucked into pants, or occluded regions when the model is sideways. Evaluate whether the model can recover visually continuous and coherent structures, rather than producing blur or distortion. Note: It does not need to be completely consistent with the ground truth. Method:Method AMethod BMethod CMethod DMethod E Score:n 0n 1n 2n 0n 1n 2n 0n 1n 2n 0n 1n 2n 0n 1n 2 Q4: Overall Preference Description: Which method's result do you prefer overall? Hint: Please select your favorite result based on your intuitive feelings. You don't need to consider any specific criteria, just choose based on your overall impression. n Method An Method Bn Method Cn Method Dn Method E Fig. 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. BridgeDiff for Virtual Try-Off31 Input Target BridgeDiff (Ours) MGT Any2AnyTryon Fig. 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. 32S. Liu et al. Input Target BridgeDiff (Ours) MGT Any2AnyTryon Fig. 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. BridgeDiff for Virtual Try-Off33 Input TryOffAnyone TryOffDiff MGT Any2AnyTryon Target BridgeDiff (Ours) Fig. 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.