acceptodds
Under review as a conference paper at ICLR 2027

SymFashion: Learning Cross-Task Symmetry for Unified Fashion Generation

Abstract

Can compatible forward and reverse fashion transformations provide useful supervision for one another, or does joint training merely share parameters? We address this question by proposing SymFashion, a unified fashion-generation model built on a single rectified-flow transformer that supports cloth-to-person try-on (C2P), person-to-person (P2P) garment transfer, virtual Try-Off, and pose-controllable cloth-to-model (C2M) generation. SymFashion maps compatible directions to a shared human–garment state and explicitly couples their role-aligned clean endpoints at the same timestep using permutation-aligned noise. Accounting for the latent mismatch introduced by the two encoding paths, our analysis recovers -weighted velocity agreement under exact latent alignment and supplies a descriptive cross-view root-risk bound. On VITON-HD and DressCode, a single checkpoint remains competitive across all four task families. Under a fixed total optimization budget, uncoupled joint training has mixed effects on the primary task-level readouts: it degrades C2P and C2M but improves P2P and Try-Off, whereas symmetric coupling improves both regimes. Direct diagnostics reveal substantially lower cross-view disagreement, especially at higher noise levels and in person-side garment regions, despite higher branchwise prediction risk. Thus, compatible transformations can support one another, but the benefit is more consistent across our evaluations when their shared semantic structure is explicitly coupled rather than left to uncoupled joint training.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Virtual Tryon, Symmetric Fashion

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

SymFashion uses a FLUX-based model for cloth-to-person generation, person-to-person garment transfer, try-off and cloth-to-model generation. Each task is represented on a two-slot person/garment canvas. During training, compatible forward and reverse views share a timestep and aligned noise, and a garment-weighted loss aligns their predicted clean endpoints after role permutation. The paper analyzes this consistency term while accounting for differences between joint and separate VAE encoding. Experiments on VITON-HD and DressCode compare the coupled model with the same model trained without the consistency loss, alongside specialist baselines and ablations of the coupling design.

Critical strengths
  1. The two-slot representation gives the four tasks a common structure and makes the proposed consistency loss straightforward to apply. Compatible views come from the same person-garment sample, so the constraint has a concrete relationship to the desired output. This is a practical way to share information between tasks with different input and output roles.

  2. The coupled and uncoupled models use the same architecture, data, supervised views and training budget. The coupled model improves VITON-HD unpaired C2P FID from 9.39 to 8.89, Try-Off FID from 8.55 to 8.20, and no-pose C2M mean-pairwise LPIPS from 0.252 to 0.226. The noise, weighting, gradient-flow and endpoint-loss ablations help explain which implementation choices matter.

  3. The analysis accounts for the VAE encoding residual and examines what the consistency loss actually changes. Appendix A reports both cross-view disagreement and native denoising error, including the increase in the latter. Reporting both quantities makes the mechanism easier to understand and gives readers more than an aggregate image-quality table.

Critical weaknesses
  1. Genuine P2P transfer is evaluated mainly with FID and KID, which do not measure whether the output preserves the specified source garment. The paired P2P result uses the unmasked target as the reference and measures self-reference reconstruction. A direct garment-fidelity evaluation is needed to judge the transfer improvement, particularly for patterns, logos and details that distribution-level metrics can miss.

  2. Several image-level gains are small, such as unpaired P2P FID 9.60 to 9.37 and KID 1.31 to 1.27 on VITON-HD. The bootstrap analysis covers latent diagnostics for selected checkpoints, while the uncertainty of the main image-quality differences is not reported. This makes it difficult to judge how consistent the benefit is across evaluation samples and training runs.

  3. Performance is sensitive to the strength of the coupling loss. In Table 4, increasing λ from 1 to 2 worsens several metrics and can erase the improvement over the uncoupled model. The paper does not explain clearly how λ=1 was selected or whether the same choice remains effective when the task mixture changes. That matters for using the method beyond the reported configuration.

Questions for authors
  1. On genuine unpaired P2P transfer, does coupling improve garment identity and detail preservation over the matched uncoupled model? Can you provide a source-conditioned metric or a blinded pairwise assessment?

  2. Can you report uncertainty for the matched image-level gains, particularly the small P2P differences? Please distinguish variation from resampling evaluation data from variation across independent training runs, where available.

  3. How was the coupling weight chosen, using which validation set? Does one fixed value work across the task mixtures in the paper, or does each mixture require retuning?

6 comments
AuthorsOpenReview ↗

Summary and Discussion Note

Thank you for your constructive feedback. In response, we added direct source-conditioned evaluation for genuine unpaired P2P, reported evaluation-sample uncertainty for the new source-conditioned metrics, added an independent retraining check, and expanded the analysis of the coupling-weight sensitivity and selection.

These additions have been incorporated into the revised manuscript.

Discussion note. All figure, table, equation, and line references in the responses below refer to the revised manuscript. Newly added or modified content is highlighted in orange in the revised manuscript, while unchanged content retains its original formatting. Because new figures and tables were inserted during revision, some pre-existing figures and tables were automatically renumbered relative to the submitted manuscript. Fig. 2 was moved from the appendix to the main paper for clearer presentation; its content was already present in the submitted manuscript.

Please find our detailed responses below.

AuthorsOpenReview ↗

W1 + Q1: Genuine unpaired P2P garment fidelity

W1. Genuine P2P transfer is evaluated mainly with FID and KID, which do not measure whether the output preserves the specified source garment. The paired P2P result uses the unmasked target as the reference and measures self-reference reconstruction. A direct garment-fidelity evaluation is needed to judge the transfer improvement, particularly for patterns, logos and details that distribution-level metrics can miss.

Q1. On genuine unpaired P2P transfer, does coupling improve garment identity and detail preservation over the matched uncoupled model? Can you provide a source-conditioned metric or blinded pairwise assessment?

Response

Thank you for this comment. Yes, on genuine unpaired P2P transfer, symmetric coupling improves source-garment consistency over the matched uncoupled model.

We added Mask-DINO and CLIP-I for genuine unpaired P2P in Table 4(a), Lines 385–394. The metric definition and paired-bootstrap uncertainty protocol are described in Lines 397–401. Mask-DINO is the cosine similarity between DINOv2 features pooled over Sapiens-derived garment regions in the source and generated output:

Train→eval.Metric ↑w/o sym.w/ sym.Δ [95% CI]n
512→512Mask-DINO0.82920.8382+0.0090 [+0.0021, +0.0158]2030
512→512CLIP-I0.92490.9287+0.0039 [+0.0012, +0.0066]2030
512→1024×1024Mask-DINO0.78810.8060+0.0179 [+0.0099, +0.0260]2028
512→1024×1024CLIP-I0.89940.9065+0.0071 [+0.0039, +0.0104]2028

All paired 95% CIs exclude zero.

Appendix Fig. C provides complementary qualitative evidence: across the displayed P2P examples, the symmetric variants more consistently preserve the source garment’s style and identity, with fewer visible garment-structure and outfit-drift artifacts.

AuthorsOpenReview ↗

W2 + Q2: Magnitude of the P2P gain and uncertainty

W2. Several image-level gains are small, such as unpaired P2P FID 9.60→9.37 and KID 1.31→1.27 on VITON-HD. The bootstrap analysis covers latent diagnostics for selected checkpoints, while the uncertainty of the main image-quality differences is not reported, making it difficult to judge consistency across evaluation samples and training runs.

Q2. Can you report uncertainty for the matched image-level gains, particularly the small P2P differences? Please distinguish evaluation resampling from independent training runs.

Response

Thank you for this comment. We address the magnitude of the matched P2P gain and the scope of the reported uncertainty separately below.

(a) Magnitude of the P2P gain

The matched P2P improvement is meaningful in the context of recent VTON studies. Our 9.60→9.37 FID ((Δ=0.23\Delta = 0.23)) improvement on VITON-HD is comparable to the 0.254 unpaired-FID margin reported by DeCo-VTON (ECCV 2026) and larger than the 0.013 margin reported by CCVTON (CVPR 2026). For Voost (SIGGRAPH Asia 2025), the final model is 0.159 FID better than the second-best external method on VITON-HD (9.141→8.982), while its own w/o λ-scaling ablation differs by only 0.009 FID (8.991→8.982). RefTON (CVPR 2026) reports internal gains of 0.15 FID on VITON-HD (8.58→8.43) and 0.24 FID on DressCode (5.31→5.07). On DressCode, our 4.59→4.38 FID ((Δ=0.21\Delta = 0.21)) improvement is also comparable to DeCo-VTON’s 0.170 margin. These values are provided only to contextualize the scale of FID differences commonly treated as meaningful in recent VTON studies, rather than as direct method-to-method comparisons.

The matched P2P results are:

MethodVITON-HD FID ↓VITON-HD KID ×10³ ↓DressCode FID ↓DressCode KID ×10³ ↓
Ours w/o sym.9.601.314.590.85
Ours w/ sym.9.371.274.380.60

(b) Evaluation-sample uncertainty and independent retraining

As discussed above, the magnitude of the matched P2P gains is comparable to differences reported and treated as meaningful in recent VTON studies.

Importantly, FID and KID are distribution-level metrics computed over sets of generated and reference images, rather than per-sample measurements. Therefore, the paired per-example uncertainty analysis used for our source-conditioned metrics does not directly apply to the reported FID/KID differences.

For the new Mask-DINO and CLIP-I evaluations, each matched example yields an individual source-conditioned similarity score. This enables paired-bootstrap resampling over the same test samples, from which we report 95% confidence intervals:

Train→eval.Metric ↑w/o sym.w/ sym.Δ [95% CI]n
512→512Mask-DINO0.82920.8382+0.0090 [+0.0021, +0.0158]2030
512→512CLIP-I0.92490.9287+0.0039 [+0.0012, +0.0066]2030
512→1024×1024Mask-DINO0.78810.8060+0.0179 [+0.0099, +0.0260]2028
512→1024×1024CLIP-I0.89940.9065+0.0071 [+0.0039, +0.0104]2028

All four source-conditioned gains exclude zero under paired bootstrap resampling, quantifying evaluation-sample variability.

As requested by the reviewer, and given the limited rebuttal period, we additionally performed an independent retraining with a different random seed on a different machine, using the same training code and configuration, to verify whether the observed P2P improvement reproduces under a separate training execution. On VITON-HD P2P, the independently trained uncoupled and coupled models obtain 9.79 and 9.40 FID, respectively, compared with 9.60 and 9.37 FID in the original matched run. The improvement therefore reproduces in the additional run, with coupling improving P2P FID by 0.39.

Together with the magnitude analysis above, this additional run provides further evidence that the observed P2P gain is consistent and not specific to a single training execution.

We clarified the scope of the paired-bootstrap uncertainty analysis in Lines 400–402.

Reviewer_hUV4OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether compatible fashion generation tasks can mutually benefit through a mechanism termed “symmetric training/coupling,” rather than through parameter sharing alone. Experiments are conducted on the VITON-HD and DressCode benchmarks.

Critical strengths
  1. The proposed method unifies multiple fashion synthesis tasks within a single model by exploiting the compatibility among tasks.

  2. Compared with the evaluated baselines, the proposed method achieves comparable or better performance on VITON-HD and DressCode.

Critical weaknesses
  1. Clarity and writing quality

The manuscript is difficult to follow, and several key terms are insufficiently defined. I found substantial portions of the paper hard to understand. For example:

(1) What does “maps their complete target states to common person–garment coordinates” mean? (Line 063)

(2) What does “a cross-view root-risk upper bound” mean? (Line 092)

(3) What do “Dense Person” and “Dense Garment” mean? (Line 211)

(4) What does “higher branchwise risk” mean? (Line 482)

(5) Why is a semicolon used in “This describes the induced behavior; sampledimage results establish its task-level benefit.”? (Line 482)

  1. Figures

The figures are also of limited quality and are not sufficiently self-explanatory.

(1) The lower part of Fig. 1 has low resolution.

(2) Line 156 states that “Fig. 2 summarizes two stages.” However, Panel (A) of Fig. 2 only explains the different conditions for four tasks. Panel (B) of Fig. 2 appears to contradict the statement “Predictions are mapped back to a common P–G ordering before comparison.” (Line 188), and it also makes the symmetric training resemble a reconstruction task without generating a new task-specific visual output.

(3) The titles of the two subfigures in Fig. 5 are awkward.

  1. Additional concerns about the proposed method

If I understand correctly, the proposed method has several additional limitations:

(1) Condition concatenation on a single canvas is already a common practice in existing VTON methods and appears conceptually similar to the proposed “shared two-slot latent.”

(2) The proposed method may not scale across different fashion datasets because, under the definition of “compatibility,” “the paired views must refer to the same underlying human–garment example” (Line 205).

(3) The role-aligned clean-endpoint consistency requires full denoising during training in order to derive the clean-endpoint estimates.

  1. Reference errors

(1) The author list for “CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models, ICLR’25” is incorrect. It is missing “Shiyue Zhang” and “Xujie Zhang.”

(2) The year for “OmniTry: Virtual Try-On Anything without Masks” is incorrect; it should be 2025 rather than 2026.

Questions for authors

See weaknesses. Answering all the necessary questions to make the proposed method more clear.

5 comments
AuthorsOpenReview ↗

Summary and Discussion Note

We appreciate the opportunity to clarify several aspects of the method. In the revision, we make the common [P | G] ordering, the distinction between task-specific condition construction and symmetric training, the interpretation of the diagnostic terms, and the clean-endpoint computation more explicit. We also revised the relevant figures, terminology, and reference entries.

Discussion note. All figure, table, equation, and line references in the responses below refer to the revised manuscript. Newly added or modified content is highlighted in orange in the revised manuscript, while unchanged content retains its original formatting. Because new figures and tables were inserted during revision, some pre-existing figures and tables were automatically renumbered relative to the submitted manuscript. Fig. 2 was moved from the appendix to the main paper for clearer presentation; its content was already present in the submitted manuscript.

Please find our point-by-point responses to W1–W4 below.

AuthorsOpenReview ↗

W1: Clarity and writing

W1. What does “maps their complete target states to common person–garment coordinates” mean? (Line 063)

What does “a cross-view root-risk upper bound” mean? (Line 092)

What do “Dense Person” and “Dense Garment” mean? (Line 211)

What does “higher branchwise risk” mean? (Line 482)

Why is a semicolon used in “This describes the induced behavior; sampled-image results establish its task-level benefit.”? (Line 482)

Response. We address each terminology and writing point separately below for clarity.

(a) Common person–garment coordinates

Thank you for asking for clarification. “Common person–garment coordinates” refers to mapping both predictions into the same semantic [P | G] ordering, where P denotes the person-side role and G denotes the garment-bearing role, before applying the cross-view comparison.

We have revised the corresponding description to state this semantic ordering more directly; see Lines 177–190.

(b) Cross-view root-risk upper bound

Thank you for asking for clarification. The cross-view root-risk upper bound is a diagnostic bound on the forward-view root risk, where root risk is the square root of the expected squared clean-endpoint error, i.e., a latent RMS error.

The two views are first aligned to the same semantic [P | G] ordering. Let z0⋆z_0^\star denote the evaluator reference latent, constructed from the person-side and garment-side VAE encodings after role alignment. We define:

  • Rfx0R_f^{x_0}: the forward-view squared clean-endpoint error relative to z0⋆z_0^\star;
  • Rrx0R_r^{x_0}: the reverse-view squared clean-endpoint error relative to z0⋆z_0^\star;
  • Cx0C^{x_0}: the cross-view disagreement between the aligned forward and reverse clean-endpoint predictions.

Specifically,

R_r^{x_0}=\mathbb{E}\|\bar{\hat z}_0^r-z_0^\star\|_W^2,\quad C^{x_0}=\mathbb{E}\|\hat z_0^f-\bar{\hat z}_0^r\|_W^2.$$ By Minkowski’s inequality, $$\sqrt{R_f^{x_0}} \leq \sqrt{R_r^{x_0}}+\sqrt{C^{x_0}} =B.$$ Thus, $B$ upper-bounds the **forward-view root risk** using the reverse-view estimation error and cross-view disagreement. **It is used only as a descriptive diagnostic, not as a perceptual-quality or generalization guarantee.** This interpretation was already clarified in **Appendix G, Eq. 30, Lines 1674–1681**. We now state it more directly in **Lines 309–312** for easier reading. ### (c) Dense Person and Dense Garment Thank you for asking for clarification. **“Dense Person” and “Dense Garment” denote the DensePose conditions associated with the Person and Garment roles, respectively.** They are conditioning inputs rather than additional generated targets. We have clarified this terminology in **Lines 179–180**. ### (d) Higher branchwise risk Thank you for asking for clarification. **“Higher branchwise risk” means higher individual clean-latent prediction error for each branch.** This refers to a single-timestep latent diagnostic and is **not** a statement about FID, KID, or sampled-image quality. We have clarified this terminology in **Line 485**. ### (e) Semicolon sentence Thank you for pointing this out. We revised the sentence to: **“These diagnostics characterize the behavior induced by coupling. Separately, sampled-image evaluations establish its task-level benefit.”** See **Lines 535–537**.
AuthorsOpenReview ↗

W2: Figures

W2. The lower part of Fig. 1 has low resolution.

Line 156 states that “Fig. 2 summarizes two stages.” However, Panel (A) of Fig. 2 only explains the different conditions for four tasks.

Panel (B) of Fig. 2 appears to contradict the statement “Predictions are mapped back to a common P–G ordering before comparison.” (Line 188)

It also makes the symmetric training resemble a reconstruction task without generating a new task-specific visual output.

The titles of the two subfigures in Fig. 5 are awkward.

Response. We address the figure-related comments separately below for clarity.

(a) Fig. 1 resolution

The figure was downsampled during PDF compression to meet the submission size limit, which reduced its visual quality; however, the content remains legible in the submitted PDF.

(b) Fig. 2 stages

Thank you for this comment. Panel (A) defines the task-specific condition construction for C2P, P2P, C2M, and Try-Off, while Panel (B) presents the symmetric training procedure. Panel (A) is therefore not a separate generation stage.

To avoid ambiguity, we revised the original “two stages” wording accordingly in Fig. 3 and Lines 171–180.

(c) Common P–G ordering in Panel (B)

In Panel (B), the reverse-view endpoint is mapped by π−1\pi^{-1} back to the forward semantic [P | G] ordering before the consistency loss is applied. The comparison is therefore performed between role-aligned clean-endpoint predictions, rather than between the original physical tile positions.

We clarified this mapping in Fig. 3(B) and Lines 175–180.

(d) Reconstruction-task interpretation

The symmetric training is not a reconstruction task or sequential reconstruction cycle. The two branches remain task-specific predictions supervised by their own flow-matching objectives. At the same sampled timestep, SymLoss compares their role-aligned clean-endpoint estimates after mapping them to the same semantic [P | G] ordering, using the garment-mask-weighted L2L_2 consistency loss.

Therefore, no generated image from one task is used as the reconstruction target or input of the other task. The two branches are evaluated in parallel at the same timestep and are coupled only through their aligned clean-endpoint predictions.

We clarified this distinction in Lines 285–289; Eqs. 8–10 provide the corresponding consistency and supervised training objectives.

(e) Subfigure titles

We revised the two subfigure titles for clarity to “Spatial disagreement by role” and “Person-role regional disagreement”; see Fig. 7.

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

SymFashion unifies four fashion generation tasks—virtual try-on, person-to-person garment transfer, virtual try-off, and garment-conditioned human generation—in a single model. It studies how these tasks interact during joint training and introduces a role-aligned consistency loss to encourage agreement between compatible views of the same person–garment pair. Experiments and ablation studies on VITON-HD and DressCode evaluate the benefits of this constraint over standard joint training.

Critical strengths
  1. The framework is clear and straightforward, supporting four fashion generation tasks with a single model. The role-aligned consistency objective provides a simple way to connect related tasks during training.

  2. The controlled comparisons and ablation studies are fairly thorough. They help isolate the effect of the proposed consistency loss from standard joint training and clarify the contributions of its main components.

Critical weaknesses
  1. The methodological novelty appears limited relative to prior work on unified fashion generation and cross-task consistency, including Voost and CCVTON. The main addition is the role-aligned parallel endpoint constraint, but its advantage over closely related consistency-based approaches is not sufficiently established.

  2. All experiments are evaluated at 512×512, with baseline outputs also resized to this resolution. This may obscure differences in fine textures, printed patterns, and garment boundaries. The current results therefore do not establish whether the reported benefits extend to high-resolution generation and garment-detail preservation.

Questions for authors
  1. What is the key methodological advance over existing cross-task and diffusion consistency methods, beyond adapting them to aligned person–garment representations?

  2. Could you provide higher-resolution comparisons with and without the proposed consistency loss to verify whether it improves the preservation of fine garment details, such as textures, printed patterns, and boundaries?

6 comments
AuthorsOpenReview ↗

Summary and Discussion Note

Thank you for your constructive feedback! To address your concerns, we clarified the methodological distinction from prior cross-task approaches and added matched zero-shot 1024×1024 evaluations. These additions have been incorporated into the revised manuscript.

Discussion note. All figure, table, equation, and line references in the responses below refer to the revised manuscript. Newly added or modified content is highlighted in orange in the revised manuscript, while unchanged content retains its original formatting.

Because new figures and tables were inserted during revision, some pre-existing figures and tables were automatically renumbered relative to the submitted manuscript. Fig. 2 was moved from the appendix to the main paper for clearer presentation; its content was already present in the submitted manuscript.

Please find our detailed responses below.

AuthorsOpenReview ↗

W1 + Q1: Methodological novelty and distinction from prior work

W1. Methodological novelty appears limited relative to prior unified fashion generation and cross-task consistency methods, including Voost and CCVTON; the advantage of role-aligned parallel endpoint coupling is not sufficiently established.

Q1. What is the key methodological advance over existing cross-task and diffusion consistency methods, beyond adapting them to aligned person-garment representations?

Response

We appreciate the reviewer’s request to further clarify how SymFashion differs from prior unified fashion generation and cross-task consistency approaches. In the revision, we make these distinctions more explicit and provide additional evidence on the effectiveness of the proposed coupling design.

(a) SymFashion evaluates compatible task directions in parallel at the same timestep with permutation-aligned noise, maps their predictions to a common [P | G] ordering, and directly couples their predicted clean endpoints. This differs from Voost, which uses optional inference-time self-correction, and CCVTON, which sequentially composes Try-Off and Try-On through generated intermediate images, with flow/DISTS losses for person reconstruction and VGG-based garment regularization. In contrast, SymFashion imposes parallel prediction-level consistency directly between role-aligned clean endpoints, without generated intermediate images or auxiliary-task inference.

(b) To evaluate the effectiveness of this design, Table 5 compares the proposed clean-endpoint coupling with the matched uncoupled model and alternative prediction-space consistency objectives under the same task configuration. Clean-endpoint coupling achieves 8.89/9.37/8.20/0.226 on C2P/P2P/Try-Off/C2M, compared with 9.39/9.60/8.55/0.252 without symmetry, 10.78/11.02/9.66/0.242 with raw velocity consistency, and 10.04/10.29/8.96/0.229 with (\sigma^2)-weighted velocity consistency. These results show that the benefit is not explained by the shared representation alone and support the effectiveness of the proposed clean-endpoint coupling over the tested consistency alternatives.

Following the reviewer’s comment, we clarified these methodological distinctions in Fig. 2 and Lines 150–154, with the corresponding task construction and semantic alignment detailed in Fig. 3 and Lines 171–180. Table 5 provides the corresponding prediction-space comparison between clean-endpoint coupling, the uncoupled model, and the tested velocity-space alternatives.

AuthorsOpenReview ↗

W2 + Q2: High-resolution behavior and fine garment details

W2. All experiments are evaluated at 512×512; this may obscure fine textures, printed patterns, and garment boundaries, and does not establish whether the benefits extend to high resolution.

Q2. Could you provide higher-resolution comparisons with and without the proposed consistency loss to verify preservation of fine garment details?

Response

Thank you for this suggestion! We agree that higher-resolution evaluation provides a useful complementary assessment of the method’s performance beyond the 512×512 setting. We therefore added a matched zero-shot 1024×1024 evaluation using the same 512-trained checkpoints.

The newly added Table 4 summarizes both the source-conditioned P2P evaluation and the matched zero-shot 1024×1024 evaluation:

Table 4(a): VITON-HD unpaired-P2P source-garment consistency

Train→eval.Metric ↑w/o sym.w/ sym.Δ [95% CI]n
512→1024×1024Mask-DINO0.78810.8060+0.0179 [+0.0099, +0.0260]2028
512→1024×1024CLIP-I0.89940.9065+0.0071 [+0.0039, +0.0104]2028

Table 4(b): VITON-HD zero-shot 1024×1024 across four tasks

VariantC2P FID ↓P2P FID ↓Try-Off FID ↓C2M MP-LPIPS ↓
w/o sym.9.789.8118.360.290
w/ sym.8.898.7413.100.248

For Table 4(a), Mask-DINO measures the cosine similarity between DINOv2 features pooled over Sapiens-derived garment regions in the source image and the generated output, while CLIP-I provides a complementary image-level similarity measure. The paired 95% confidence intervals exclude zero.

Table 4(b) shows that symmetric coupling improves all four evaluated task readouts at 1024×1024. Fig. 4 further shows more stable zero-shot high-resolution behavior with coupling: the uncoupled model exhibits noticeable garment-style, color, or task-output errors, while coupling better preserves garment appearance and task semantics.

Together, these results show that the benefit of symmetric coupling persists beyond the 512×512 evaluation setting.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Author Response to the Area Chair and Reviewers

Dear AC and reviewers,

We sincerely thank the AC and all reviewers for their time, effort, and constructive feedback on our manuscript!

In sum, SymFashion introduces cross-task clean-endpoint supervision for unified fashion generation, where compatible task directions are aligned to a common Person–Garment representation and jointly supervised at their predicted clean endpoints.

Based on your feedback, we revised the manuscript with the following updates. We also clarified key points regarding our methodological contribution, training procedure, and evaluation:

  1. We clarified the methodological contribution and training procedure. We now more clearly distinguish SymFashion from prior cross-task approaches, explain the common [P | G] alignment and compatible-pair construction, and clarify that symmetric training is not a reconstruction cycle and does not require full denoising. (Reviewers vHLs, hUV4)
  2. We added direct source-conditioned evaluation for genuine unpaired P2P transfer. We report Mask-DINO and CLIP-I, both of which improve with symmetric coupling and have paired 95% confidence intervals excluding zero. Importantly, the matched P2P improvements (0.23 FID on VITON-HD and 0.21 FID on DressCode) are meaningful in the context of recent VTON benchmarks, where competitive methods and ablation variants often differ by comparable or smaller margins. The P2P improvement also reproduces in an independent retraining run. (Reviewer kxzB)
  3. We added zero-shot 1024×1024 evaluation using the same 512-trained checkpoints. Symmetric coupling improves all four evaluated tasks (C2P, P2P, Try-Off, and C2M) and also improves source-conditioned P2P garment consistency at the higher resolution. (Reviewer vHLs)
  4. We strengthened the analysis of the symmetric coupling objective. The revised ablations compare clean-endpoint consistency against the uncoupled model and velocity-space alternatives, and clarify the sensitivity and selection of λs\lambda_s. A single λs=1\lambda_s=1 is used across all reported task configurations and datasets without task- or dataset-specific retuning. (Reviewer kxzB)
  5. We improved the manuscript clarity and reproducibility. We clarified the interpretation of the reported uncertainty and diagnostic terminology, revised figures and evaluation descriptions, expanded implementation details, and corrected the identified reference errors. (Reviewers hUV4, kxzB)

We again thank the AC and reviewers for the constructive feedback, which significantly improved the clarity and empirical support of the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.