acceptodds
Under review as a conference paper at ICLR 2027

Looking Below the Latent Token: Detail Retrieval via Correspondence Supervision for Virtual Try-On

10 views

Abstract

Latent-space virtual try-on (VTON) models often distort fine-grained garment details, such as small text, logos, stitching, and repetitive patterns. While existing methods improve garment-person correspondence, such correspondence is defined over spatially coarse latent tokens and therefore provides only coarse localization of reference garment content. We introduce DECOR-VTON, a two-phase framework that uses latent correspondence as a spatial routing signal for selectively accessing finer garment information. Through layer-wise analysis and controlled attention interventions, we identify late dual-stream blocks as promising correspondence-aware representations. Phase I improves correspondence in these blocks using probabilistic warp-consistency supervision without requiring dense correspondence annotations or an external correspondence teacher during training. In Phase II, the refined correspondence guides the selection of relevant garment regions, from which higher-resolution convolutional features are retrieved through local cross-attention and injected into downstream transformer blocks via a timestep-conditioned residual gate. This coarse-to-fine design complements latent-space generation with selectively accessed fine-grained reference features without introducing global high-resolution attention. Experiments on VITON-HD and DressCode show consistent improvements in overall synthesis quality, while qualitative comparisons indicate better preservation of fine-grained garment structures. Extensive ablations further demonstrate the complementary roles of correspondence supervision and high-resolution detail retrieval.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Virtual Try-On, Generative Models, Coarse-to-Fine Retrieval

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xGD6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DECOR-VTON, a two-stage virtual try-on framework that leverages FLUX.1 Fill as its base model. To improve garment-to-person spatial correspondence, the first stage introduces probabilistic warp consistency, integrated into the late dual-stream attention blocks. Building on this learned correspondence, the second stage retrieves local high-resolution garment features and incorporates them into downstream single-stream blocks. The authors evaluate the proposed method on standard benchmarks (VITON-HD and DressCode) and provide ablation studies that validate the effectiveness of the supervision and feature retrieval modules.

Critical strengths
  1. The coarse-to-fine retrieval architecture is highly intuitive and well-justified. Since spatial localization alone cannot restore fine-grained information lost in coarse garment tokens, the retrieval branch effectively supplies high-resolution reference features.

  2. This method offers a practical and scalable source of self-supervision in the absence of ground-truth correspondence annotations, successfully mitigating reliance on external matching teachers or pseudo-labels.

Critical weaknesses
  1. While the qualitative examples are compelling, the global metrics on the whole image fail to show how consistently the method preserves fine details—such as text, logos, or intricate repeated patterns—across the full test set. Specifically, in Table 2, Phase I substantially improves PCK from 0.885 to 0.980, yet global metrics like SSIM and LPIPS remain unchanged at the reported precision, leaving a gap between improved correspondence and tangible detail reproduction. Furthermore, the ablation study in Table 8 conflates spatial resolution with the VGG feature layer level, making it difficult to determine whether the observed performance gains stem directly from higher-resolution feature retrieval or from layer-specific semantic representation.

  2. While the ablation study shows that combining both stages yields performance gains, it does not isolate the precise contribution of the proposed routing mechanism. For rigorous control, an experiment should freeze the garment features and injection modules and vary only the retrieved spatial regions. Without this baseline, it is unclear whether the improvements stem from accurate, correspondence-guided feature selection or merely from the capacity added by an extra feature branch. Additionally, the lack of a direct empirical comparison with CORAL—a closely related method highlighted in the related work—limits the assessment of the proposed correspondence supervision mechanism.

Questions for authors
  1. Could the authors provide a targeted quantitative or qualitative evaluation focused on garments featuring text, logos, or complex repeated patterns? Additionally, to isolate the routing mechanism's efficacy, could you include a control experiment in which the retrieved regions are randomly shuffled or misaligned, while keeping the feature injection modules fixed? This would help clarify whether the detail fidelity stems from accurate correspondence-guided routing rather than simply increasing feature capacity.

  2. Could the authors clarify the exact implementation details for computing PCK@16? Please clarify: 1) the selection criteria for evaluation keypoints and any pseudo-label filtering mechanisms applied; 2) the specific attention layers and diffusion timesteps used for feature extraction; 3) whether the person features are extracted from noisy ground-truth images or from generated intermediate samples. Furthermore, is a direct empirical comparison with CORAL supervision available under the same unified training setup?

7 comments
AuthorsOpenReview ↗

Official Author Responses to Reviewer xGD6 [1/7]

We sincerely thank Reviewer xGD6 for recognizing the intuitive and well-justified coarse-to-fine retrieval architecture and its teacher-free correspondence supervision. We appreciate the requests for more direct evidence of fine-detail preservation and causal routing. Below, we address (1) automatically defined garment-attribute subsets drawn from the full test split, (2) a fixed-resolution VGG abstraction-level ablation, (3) shuffled-region routing with unchanged model weights, (4) the PCK evaluation protocol and stricter thresholds, and (5) controlled comparisons with CORAL. The revised paper includes these analyses in Tables 1–2 and 14–17 and Appendices B.6 and C.9–C.12.

AuthorsOpenReview ↗

Official Author Responses to Reviewer xGD6 [2/7]

W1.1 + Q1.1: Detail Fidelity on Text, Logos, and Repeated Patterns

W1.1 The global metrics on the whole image fail to show how consistently the method preserves fine details—such as text, logos, or intricate repeated patterns—across the full test set. Specifically, in Table 2, Phase I substantially improves PCK from 0.885 to 0.980, yet global metrics like SSIM and LPIPS remain unchanged at the reported precision, leaving a gap between improved correspondence and tangible detail reproduction.

Q1.1 Could the authors provide a targeted quantitative or qualitative evaluation focused on garments featuring text, logos, or complex repeated patterns?

Unchanged SSIM and LPIPS. We thank the reviewer for this observation. We would like to emphasize that the discrepancy between correspondence accuracy and image reconstruction metrics in Table 2 is consistent with the central motivation of our two-phase framework. Our hypothesis is that improving garment–person correspondence alone is insufficient to fully preserve fine-grained garment details.

Fine-grained detail evaluation. To evaluate garment detail preservation, we automatically categorize all 2,032 VITON-HD test garments using Qwen3-VL-32B-Instruct into three overlapping groups: text/typography (1,140), logos/graphic prints (1,054), and complex repeated patterns (628). To avoid manual test-set selection, the categories are determined solely from reference garment images, independently of generated results. We then compute garment-masked spatial LPIPS between generated and ground-truth images and report the average scores for each category. DECOR-VTON achieves the lowest garment-region LPIPS across all three categories, with approximately 8.8%, 8.7%, and 7.7% relative reductions over DECOR-VTON (Baseline), respectively, based on the rounded scores. It also outperforms CORAL, Leffa, and OrthoTryOn across these subsets, providing quantitative evidence of improved garment fidelity on images containing fine-grained details (revised Appendix C.12, Table 17).

Table R5. Garment-masked spatial LPIPS (↓) on paired VITON-HD, stratified by automatically identified garment attributes. Best results are shown in bold.

MethodText (1,140) ↓Logos (1,054) ↓Patterns (628) ↓
CORAL0.14200.14110.1600
Leffa0.14630.14620.1662
OrthoTryOn0.19690.19560.2154
DECOR-VTON (Baseline)0.14480.14330.1629
DECOR-VTON (Phase I only)0.14370.14260.1638
DECOR-VTON (w/o Phase I)0.13960.13830.1577
DECOR-VTON (Ours)0.13200.13080.1504

Key takeaway: On overlapping attribute subsets selected from all 2,032 VITON-HD test garments, DECOR-VTON achieves the lowest garment-masked LPIPS for text (0.1320), logos (0.1308), and complex patterns (0.1504), without selecting examples based on generated outputs.

AuthorsOpenReview ↗

Official Author Responses to Reviewer xGD6 [3/7]

W1.2: Separating VGG Feature Resolution from Abstraction

Furthermore, the ablation study in Table 8 conflates spatial resolution with the VGG feature layer level, making it difficult to determine whether the observed performance gains stem directly from higher-resolution feature retrieval or from layer-specific semantic representation.

Table R3. Controlled ablation of VGG feature abstraction levels at a fixed spatial resolution on VITON-HD (revised-paper Table 14). Best results are shown in bold.

MethodFID (P/U) ↓KID (P/U) ↓SSIM ↑LPIPS ↓DISTS ↓
VGG Block 1 (128 × 96)4.63 / 8.760.20 / 0.830.8960.0400.065
VGG Block 2 (128 × 96)4.58 / 8.810.19 / 0.850.8950.0400.066
VGG Block 3 (128 × 96)4.58 / 8.750.21 / 0.760.8950.0400.065

Effect of retrieval feature resolution and abstraction level. Table 8 evaluates features from different VGG blocks at spatial resolutions ranging from 512×384512\times384 to 64×4864\times48. To further examine the effect of feature abstraction, we conduct a controlled experiment using VGG Blocks 1–3 at the same resolution of 128×96128\times96, with identical retrieval and injection settings (Table R3; revised-paper Table 14). VGG Block 3 achieves the lowest unpaired FID/KID (8.75/0.76) and a paired FID of 4.58, matching Block 2. The comparable results across blocks suggest that different feature abstraction levels can support effective retrieval at a fixed spatial resolution.

Key takeaway: Feature abstraction is evaluated at a fixed 128×96128\times96 resolution: VGG Block 3 offers the best unpaired FID/KID, but overall performance remains comparable across Blocks 1–3.

Reviewer_YXxWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an image virtual try-on method that selectively enhances the texture of clothing details. It uses a two-stage training method. The first stage uses random Warping to enable the model to select high-resolution feature tokens related to top-k. The second stage injects the tokens into the backbone generation network. The results show that it has a certain effect.

Critical strengths

The implementation and application of the method in this paper are technically feasible and meet the scenario requirements in the field of virtual try-on. The experiments in this paper are relatively complete, the writing is clear, and the figures and tables are easy to read;

Critical weaknesses
  1. Figure 1 illustrates that the method proposed in this paper combines Latent-Space and Pixel-Space. However, the method base itself is still Latent Space, indicating that Latent Space itself does not cause detail blurring. In addition, the VGG used in this paper also extracts high-res features with a 4x downsampling, and only takes the Top-4. It also does not specify whether targeted fine-tuning has been performed or whether pre-trained VGG is directly used in training. This raises questions about whether this high-res information is sufficient.

  2. In addition, the paper claims to alleviate the High computational cost of Pixel-Space, but the full text does not provide any comparison of metrics related to computational efficiency. From an engineering perspective, in order to obtain top-k, it is necessary to completely calculate the attention matrix from person to garment in dual-stream blocks. FLUX usually uses flash attention and doesn't save this matrix, so it needs to explicitly calculate each step again before performing top-k and gather. gather operations such as cutting windows by index and stitching often appear slower on the GPU than FLOPs numbers.

  3. The first phase of training uses a random Warping method, but whether it is Homography, TPS, or Affine + TPS, it is difficult to simulate the 3D deformation of the fabric. This may make the learned deformation focus relatively monotonous and simple, which is insufficient to cope with real application scenarios.

Questions for authors
  1. Can you report the inference time, memory, and parameters at 512×384 and 1024×768 compared to the FLUX baseline (and one or two recent VTON methods)? top-k needs to explicitly calculate the attention matrix from person to garment, which flash attention typically does not save. How much additional overhead does this incur for each step?
  2. VGG-16 Is it frozen? Is it pre-trained by ImageNet? Why VGG? Have you tried others, such as DINO features, or directly using pixel patches?
  3. It is trained an extra 50K steps in Phase II. Is there a baseline that trains the same number of steps, but without the retrieval module? Only in this way can we distinguish whether the improvement comes from the retrieval itself or from longer training.
  4. Phase I only used homography, TPS, and affine+TPS. How do the learned correspondences perform when encountering real-world situations with significant 3D deformation, wrinkles, or occlusion? Could you show some failure cases?
9 comments
AuthorsOpenReview ↗

Official Author Responses to Reviewer YXxW [1/7]

We thank Reviewer YXxW for the technically grounded review and for recognizing the practical feasibility of the approach, the relatively complete experimental evaluation, and the clear writing and figures. We address the concerns about (1) the distinction between latent-space generation and fine-resolution reference retrieval, (2) VGG initialization and backbone selection, (3) explicit attention reconstruction and actual inference overhead, (4) a 120K-step training-budget-matched baseline, and (5) the scope and limitations of synthetic warp supervision. The relevant quantitative updates appear in the revised-paper Table 3 and Appendix C.8, with changes highlighted in red.

AuthorsOpenReview ↗

Official Author Responses to Reviewer YXxW [2/7]

W1.1: Motivation and Effectiveness of High-Resolution Retrieval

Figure 1 illustrates that the method proposed in this paper combines Latent-Space and Pixel-Space. However, the method base itself is still Latent Space, indicating that Latent Space itself does not cause detail blurring. In addition, the VGG used in this paper also extracts high-res features with a 4x downsampling, and only takes the Top-4. It also does not specify whether targeted fine-tuning has been performed or whether pre-trained VGG is directly used in training. This raises questions about whether this high-res information is sufficient.

Latent-space generation versus high-resolution retrieval. We agree that latent-space generation is not inherently blurry. However, latent-token correspondence may lack the granularity to capture fine details such as text, logos, and repetitive patterns. DECOR-VTON therefore supplements the latent-space backbone with higher-resolution garment features to mitigate detail loss, while acknowledging that retrieval cannot fully recover information lost during VAE compression, as discussed in Appendix E.

Feature resolution and Top-k selection. Although VGG features are downsampled by 4×, they retain twice the spatial resolution of latent features along each dimension. Thus, each latent token corresponds to four VGG feature patches, and Top-4 provides access to 16 fine-grained patches through local cross-attention. Our ablations identify 4× downsampling as the best-performing resolution among those evaluated, although the proposed retrieval mechanism supports higher feature resolutions. Top-4 also achieves the best unpaired FID/KID (8.73/0.78), with no further improvement from Top-8 (8.79/0.81).

Key takeaway: Latent-space generation is not inherently blurry; our contribution addresses the coarse spatial granularity of latent-token correspondence by accessing 16 higher-resolution feature locations from four selected garment regions.

AuthorsOpenReview ↗

Official Author Responses to Reviewer YXxW [3/7]

W1.2 + Q2: VGG Initialization and Encoder Selection

W1.2. Figure 1 illustrates that the method proposed in this paper combines Latent-Space and Pixel-Space. However, the method base itself is still Latent Space, indicating that Latent Space itself does not cause detail blurring. In addition, the VGG used in this paper also extracts high-res features with a 4x downsampling, and only takes the Top-4. It also does not specify whether targeted fine-tuning has been performed or whether pre-trained VGG is directly used in training. This raises questions about whether this high-res information is sufficient.

Q2. VGG-16 Is it frozen? Is it pre-trained by ImageNet? Why VGG? Have you tried others, such as DINO features, or directly using pixel patches?

VGG initialization and training. VGG-16 is initialized with ImageNet-pretrained weights and jointly fine-tuned with the retrieval modules during Phase II; it is not frozen.

Rationale for VGG-16: We select VGG-16 for its simple hierarchical architecture and high-resolution early features. Unlike standard ViT encoders, which reduce spatial resolution through patch embedding, or ResNet, which uses an initial 7×77\times7 convolution with stride 2, the early convolutional layers of VGG-16 preserve the input spatial resolution. Its hierarchical structure also enables controlled analysis of spatial resolution and feature abstraction. Moreover, VGG features are widely used in perceptual metrics such as LPIPS. Our controlled ablation (Table R3) further shows comparable performance across VGG Blocks 1–3 at a fixed spatial resolution. While alternative encoders (e.g., DINO, VAE, ResNet, and ConvNeXt) and raw pixel patches could be explored, a comprehensive backbone comparison is beyond the scope of this work. Our primary contribution is the correspondence-guided retrieval mechanism rather than feature extractor optimization.

Key takeaway: VGG-16 is ImageNet-initialized and fine-tuned during Phase II; it is selected for high-resolution early features and controllable multilevel representations, not because it has been shown superior to all alternative encoders.

Reviewer_aYW3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes DECOR-VTON, a two-phase framework for preserving garment details in latent-space virtual try-on. Through layer-wise attention analysis and controlled attention interventions, the authors identify late dual-stream DiT blocks as useful sources of garment–person correspondence. Phase I adapts probabilistic warp consistency to improve attention-based correspondence without an external correspondence teacher during training. Phase II uses this correspondence to select relevant garment regions, retrieve higher-resolution VGG features through local cross-attention, and inject the retrieved information into downstream person tokens through timestep-conditioned residual gates. Experiments on VITON-HD and DressCode, including models at two resolutions, demonstrate strong synthesis performance, while ablations examine correspondence supervision and detail retrieval.

Critical strengths
  1. Clear motivation and coherent method design. The paper separates coarse correspondence-based localization from fine appearance retrieval, with each component addressing a specific part of the detail-preservation problem. The layer-wise attention analysis and controlled interventions provide useful empirical support for the design.

  2. Substantial empirical evaluation and informative correspondence-supervision ablations. Experiments cover multiple datasets and resolutions. Comparisons of correspondence objectives and ablations of the two training phases help establish the usefulness of the proposed components.

Critical weaknesses
  1. Insufficient direct comparison with the closest related DiT correspondence method. CORAL already analyzes and supervises person–garment query–key correspondence in DiTs. Although the proposed training objective does not require an external correspondence teacher, the absence of a direct comparison makes its incremental benefit difficult to assess.

  2. Incomplete specification and validation of the proposed mechanisms. The reverse correspondence construction is ambiguous between Eq. (3) and Appendix B.4. In addition, the existing ablations and visualizations do not fully explain the dependence on VGG features or how fine-grained retrieved information affects generation, and the inference overhead of Phase II is not quantified.

Questions for authors
  1. Comparison with CORAL. CORAL already analyzes and supervises person–garment query–key correspondence in DiTs. Could the authors provide a direct comparison? Comparing the two correspondence objectives with the same Phase-II retrieval module would further help distinguish the contributions of correspondence learning and detail retrieval.

  2. Reverse correspondence in Eq. (3). Equation (3) constructs the reverse distribution from the transpose of the forward affinity:

    S⊤=KGQP⊤/d.S^\top=K_GQ_P^\top/\sqrt{d}.

    Appendix B.4 states that the reverse distribution is constructed analogously from native attention. If this means using garment queries and person keys, the affinity would instead be

    QGKP⊤/d,Q_GK_P^\top/\sqrt{d},

    which is generally not equivalent. Could the authors provide the complete reverse-probability formula, including head aggregation and normalization, and reconcile the main text with the implementation?

  3. Phase-II feature analysis. Figure 6(b) visualizes the selected retrieval regions, but does not show how fine-grained information is extracted and used within those regions. Could the authors provide interpretable visualizations of the VGG features F^=VGG⁡(G)\widehat F=\operatorname{VGG}(G) and local cross-attention, and analyze the effect of the retrieved residual RpℓR_p^\ell on the updated representation

    h~pℓ=hpℓ+αℓ(t)Rpℓ?\widetilde h_p^\ell=h_p^\ell+\alpha_\ell(t)R_p^\ell?

    For example, feature-response or projected-feature visualizations could clarify the mechanism. Could the authors also compare VGG features with alternatives such as DINO features or VAE encoder intermediate features, controlling spatial resolution and retrieval settings where feasible? These experiments would help clarify the roles of spatial resolution, visual representation, and the retrieval mechanism.

  4. Computational overhead of Phase II. Phase II adds VGG encoding, local feature gathering, and cross-attention in multiple layers. Could the authors report the incremental inference latency, peak GPU memory, and parameter count relative to the same baseline?

8 comments
AuthorsOpenReview ↗

Official Author Responses to Reviewer aYW3 [1/7]

We sincerely thank Reviewer aYW3 for the detailed, constructive review and for recognizing the clear motivation, coherent method design, informative layer-wise attention analysis, and extensive empirical evaluation and ablations.

We address each concern below. Specifically, we have added (1) direct comparisons with CORAL under matched correspondence-supervision settings and full-method evaluations, (2) clarification of the reverse correspondence formulation, (3) controlled VGG feature ablations and additional visualizations of local cross-attention and gated residual injection, and (4) a detailed analysis of inference latency, GPU memory, computational cost, and parameter overhead.

The new results are presented in Tables R1–R3, with additional analyses in Appendices B.4 and C.6–C.9. Textual revisions are highlighted in red in the revised manuscript.

AuthorsOpenReview ↗

Official Author Responses to Reviewer aYW3 [2/7]

W1 + Q1: CORAL Comparison on Matched Correspondence-Supervision Settings

W1. Insufficient direct comparison with the closest related DiT correspondence method. CORAL already analyzes and supervises person–garment query–key correspondence in DiTs. Although the proposed training objective does not require an external correspondence teacher, the absence of a direct comparison makes its incremental benefit difficult to assess.

Q1. Comparison with CORAL. CORAL (https://arxiv.org/abs/2602.17636) already analyzes and supervises person–garment query–key correspondence in DiTs. Could the authors provide a direct comparison? Comparing the two correspondence objectives with the same Phase-II retrieval module would further help distinguish the contributions of correspondence learning and detail retrieval.

Missing comparison with CORAL. We thank the reviewer for highlighting this closely related work. While both methods improve person–garment correspondence in DiTs, CORAL employs DINOv3-based correspondence distillation and entropy minimization, whereas ours adopts teacher-free probabilistic warp consistency (PWC) and further exploits learned correspondence for high-resolution feature retrieval.

We conducted two comparisons: (1) correspondence supervision on the same FLUX-VTON backbone (Table R1; revised-paper Table 2) and (2) full-method evaluation on DressCode and VITON-HD (Table R2; revised-paper Table 1).

Table R1. Comparison with CORAL correspondence supervision on VITON-HD (revised-paper Table 2). Best results are shown in bold.

MethodFID (P/U) ↓KID (P/U) ↓SSIM ↑LPIPS ↓DISTS ↓PCK@16 ↑
FLUX-VTON Baseline (70K steps)4.89 / 8.960.19 / 0.870.8900.0440.0690.885
+ CORAL Loss5.02 / 8.950.24 / 0.860.8890.0450.0700.992
+ PWC (Ours)4.81 / 8.870.13 / 0.830.8900.0440.0680.980

Comparison with CORAL correspondence supervision. In Table R1, CORAL achieves higher PCK@16 (0.992 vs. 0.980), measured against DINOv3-derived pseudo-ground truth, which CORAL also uses for supervision. Our teacher-free PWC improves FID, KID, and DISTS over CORAL, while matching SSIM and LPIPS, including paired FID (4.81 vs. 5.02) and KID (0.13 vs. 0.24). In the qualitative examples of revised-paper Figure 5, CORAL exhibits sharper attention but less spatially coherent warping, illustrating that attention sharpness alone does not guarantee plausible dense correspondences.

Table R2. Full-method comparison with CORAL on DressCode and VITON-HD (revised-paper Table 1). Best results are shown in bold.

(a) DressCode

MethodFID (P/U) ↓KID (P/U) ↓SSIM ↑LPIPS ↓DISTS ↓
CORAL2.47 / 4.880.40 / 0.950.9200.0390.062
DECOR-VTON2.28 / 4.850.20 / 1.120.9260.0340.057

(b) VITON-HD

MethodFID (P/U) ↓KID (P/U) ↓SSIM ↑LPIPS ↓DISTS ↓
CORAL5.27 / 8.950.58 / 0.900.8790.0450.069
DECOR-VTON4.49 / 8.480.13 / 0.620.9040.0490.063

Full comparison with CORAL. In Table R2, at the same resolution (1024 × 768), DECOR-VTON achieves better paired FID, paired KID, SSIM, and DISTS on both datasets, along with lower unpaired FID. CORAL performs better on DressCode unpaired KID (0.95 vs. 1.12) and VITON-HD LPIPS (0.045 vs. 0.049) at this resolution.

Overall, these results show that DECOR-VTON improves the majority of reported image-generation metrics relative to CORAL across both datasets, consistent with the benefits of teacher-free correspondence supervision and high-resolution feature retrieval. The corresponding comparisons are included in Tables 1 and 2 of the revised manuscript, with textual changes highlighted in red.

Key takeaway: DECOR-VTON improves most image-generation metrics over CORAL on both datasets, while CORAL leads in DINOv3-based PCK@16, DressCode unpaired KID, and matched-resolution VITON-HD LPIPS.

AuthorsOpenReview ↗

Official Author Responses to Reviewer aYW3 [3/7]

W2.1 + Q2: Ambiguity in reverse correspondence formulas

W2.1. Incomplete specification and validation of the proposed mechanisms. The reverse correspondence construction is ambiguous between Eq. (3) and Appendix B.4. In addition, the existing ablations and visualizations do not fully explain the dependence on VGG features or how fine-grained retrieved information affects generation, and the inference overhead of Phase II is not quantified.

Q2. Reverse correspondence in Eq. (3). Equation (3) constructs the reverse distribution from the transpose of the forward affinity: Sᵀ = K_G Q_Pᵀ / √d. Appendix B.4 states that the reverse distribution is constructed analogously from native attention. If this means using garment queries and person keys, the affinity would instead be Q_G K_Pᵀ / √d, which is generally not equivalent. Could the authors provide the complete reverse-probability formula, including head aggregation and normalization, and reconcile the main text with the implementation?

Ambiguity in reverse correspondence – Equation (3) vs Appendix B.4:

We apologize for the ambiguity between Eq. (3) and Appendix B.4. Eq. (3) presents a generalized formulation applicable to both cross-attention (e.g., StableVITON) and joint-attention architectures, while Appendix B.4 describes its implementation in our joint-attention backbone.

Specifically, we first compute the person–garment affinity matrix:

SP,G=QPKG⊤dS_{P,G} = \frac{Q_P K_G^\top}{\sqrt{d}}

The forward correspondence PG←PP_{G\leftarrow P} is obtained by selecting garment probabilities from the person-query attention distribution, while the reverse correspondence PP←GP_{P\leftarrow G} is derived from the transposed affinity matrix:

SP,G⊤=KGQP⊤dS_{P,G}^\top = \frac{K_G Q_P^\top}{\sqrt{d}}

For each direction, softmax is first computed over the complete token sequence within each head, including register tokens when present. We then average probabilities across heads, retain the opposite-view entries, and renormalize over the relevant person or garment tokens (revised Appendix B.4, Eqs. 20–25). Importantly, the reverse affinity differs from native garment-to-person attention, which uses QGKP⊤Q_GK_P^\top. Since reference garment tokens are concatenated with person tokens, garment queries can directly attend to garment keys, potentially making native garment-to-person attention less reliable for reverse correspondence estimation. These details have been clarified in Appendix B.4, with changes highlighted in red.

Key takeaway: The reverse distribution uses the transpose of the person-to-garment affinity with a separate softmax and normalization over person tokens; it is not native garment-query/person-key attention.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Official Author Global Response to the AC & All Reviewers

Dear AC and reviewers,

We sincerely thank all reviewers for their careful and constructive feedback. We are encouraged that they recognized the clear motivation and well-justified coarse-to-fine retrieval design (aYW3, xGD6), the practical and technically feasible approach (YXxW, xGD6), and the self-supervised correspondence learning without external matching teachers (xGD6). We also appreciate their positive assessments of the comprehensive experiments and informative ablations (aYW3, YXxW), as well as the clear presentation and readable figures and tables (YXxW).

The shared concerns are whether the improvements come from correspondence-guided retrieval rather than additional model capacity or training, and whether the proposed mechanisms, fine-detail preservation, and computational overhead are sufficiently validated. We provide new experiments and clarifications in the revised PDF (changes highlighted in red):

  1. Direct comparisons demonstrate the effectiveness of our correspondence supervision (aYW3 W1/Q1; xGD6 W2.2/Q2.2). Under the same FLUX-VTON backbone, our teacher-free PWC achieves better paired FID (4.81 vs. 5.02) and DISTS (0.068 vs. 0.070) than CORAL supervision, although CORAL achieves higher DINOv3-based PCK@16 (0.992 vs. 0.980). The full DECOR-VTON model also outperforms CORAL on most generation metrics across VITON-HD and DressCode (Tabs. 1–2).

  2. The gains are not solely due to longer training or additional feature capacity (YXxW Q3; xGD6 W2.1/Q1.2). Against the 120K-step baseline, DECOR-VTON improves paired/unpaired FID from 4.77/8.80 to 4.61/8.73 (Tab. 3). Randomly shuffling retrieved regions with identical model weights degrades FID to 4.69/8.86, confirming the importance of correspondence-guided routing (Tab. 15).

  3. Fine-grained garment details are better preserved (xGD6 W1.1/Q1.1). We automatically identify text (1,140), logos (1,054), and repetitive patterns (628) from all 2,032 VITON-HD test garments. DECOR-VTON achieves the lowest garment-masked LPIPS across all three overlapping subsets (0.1320/0.1308/0.1504), outperforming CORAL, Leffa, and OrthoTryOn (Tab. 17).

  4. Additional analyses clarify the retrieval mechanism (aYW3 W2/Q2–Q3; xGD6 W1.2/Q2.1). We clarify reverse correspondence normalization and the PCK evaluation protocol (Apps. B.4 and B.6). Controlled VGG ablations at fixed resolution isolate feature abstraction (Tab. 14), while retrieval visualizations and gated-residual analysis reveal spatially selective attention and layer-dependent updates averaging 1.40% of the hidden-state norm (Figs. 10 and 13). We also report stricter PCK@1/2/4 evaluations (Tab. 16).

  5. Phase-II retrieval introduces modest computational overhead (aYW3 W2.3/Q4; YXxW W2/Q1). At the evaluated default configuration, inference latency, peak VRAM, FLOPs, and parameters increase by only 3.93%, 0.83%, 0.10%, and 0.14%, respectively (Fig. 11). We further clarify VGG initialization and fine-tuning (App. B.3) and discuss limitations under severe garment deformation and occlusion (App. E).

We again thank the AC and reviewers for their constructive feedback, which has substantially strengthened the clarity, reproducibility, and empirical support of our manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.