acceptodds
Under review as a conference paper at ICLR 2027

BiCorr: Learning Bidirectional Latent Correspondences for Video Virtual Try-On

Abstract

Video virtual try-on transfers a reference garment to a person in a video while preserving identity, motion, background, and garment appearance. Recent video diffusion Transformers have improved garment conditioning and temporal modeling, but fine-grained patterns and local garment structures can still be distorted, particularly under large motion and occlusion. Garment conditioning provides reference information but does not ensure that each person token retrieves relevant features from the correct garment region. To address this issue, we propose BiCorr, a framework for bidirectional latent correspondence learning between person-video and reference-garment tokens. We combine high-confidence direct matches with matches propagated along validated trajectories, and balance their coverage across time and space. These matches are mapped to the model's video and garment tokens and represented as distributional targets that account for mapping uncertainty while retaining distinct garment locations. The targets guide person tokens to retrieve features from corresponding garment regions, while a reverse objective associates garment tokens with their corresponding spatiotemporal locations throughout the video. All matching and correspondence losses are used only during training, requiring no additional matcher, tracker, or warping module at inference. Experiments show that BiCorr improves visual quality, garment fidelity, and temporal consistency.

Primary area:
generative models
Keywords:
Video Virtual Try-On, Diffusion Transformers, Latent Correspondence Learning, Attention Supervision

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8FQUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BiCorr proposes bidirectional correspondence supervision for video virtual try-on using a garment-conditioned video diffusion Transformer. It combines direct DINOv3 matches with validated tracking-based propagation, maps these frame-level correspondences to compressed latent tokens, and constructs Gaussian-based distributional targets. These targets supervise person-to-garment retrieval through self- and cross-attention, together with garment-to-person localization through self-attention. Correspondence estimation and auxiliary supervision are used only during training. The paper also introduces LHFSC, a paired spectral metric, and curates BiCorr-Test from existing official test splits. Evaluation includes video benchmarks, image try-on benchmarks, component ablations, and a 60-participant user study, with reported improvements in garment fidelity and video quality.

Critical strengths
  1. A coherent correspondence-learning approach without additional inference-time matching or tracking. The method addresses the difficulty of translating frame-level correspondences into temporally compressed video tokens and supervises both retrieval directions. The directional ablations in Table 3 show that removing either direction worsens several generation metrics, including LPIPS, LHFSC, and unpaired VFID_R. This provides meaningful evidence that bidirectional supervision contributes beyond the generation objective alone, while keeping the additional correspondence machinery confined to training.

  2. The reported improvements have quantitative and perceptual support. BiCorr performs favorably against video baselines on ViViD-S and BiCorr-Test, and the qualitative results show improved preservation of garment motifs and local structure. The user study also reports garment-detail improvements over MagicTryOn, with confidence intervals in both paired and unpaired settings. This strengthens the empirical case because it draws on established generation metrics and human judgments alongside the proposed LHFSC metric.

Critical weaknesses
  1. Insufficiently controlled ablations of the central design choices. Although Table 3 and Figure 9 provide useful quantitative and visual comparisons, several ablations change multiple factors simultaneously. The direct-only and track-only variants also remove coverage-balanced query selection and alter clip sampling. The hard-target variant changes center selection, distribution shape, and support regions; singleton supports additionally make the conditional KL term degenerate. Consequently, these experiments support the complete pipeline but do not isolate the benefits of coverage balancing, Gaussian smoothing, or retaining multiple correspondence locations. Separate ablations of these factors, with clips and other settings held fixed, are needed. The mechanism also remains incompletely explained: hard targets achieve higher mode coverage and temporal coverage than Full in Table 7 despite worse generation metrics. Finally, the ablations use a 1.3B model and 20% of the training data, whereas the main results use a 14B model. A matched generation-only comparison at the main scale would strengthen attribution of the headline improvements to correspondence supervision.

  2. Incomplete comparison with recent image-based virtual try-on methods. The image evaluation in Table 4 omits relevant recent methods, including Garments2Look, RefTon, and Mobile-VTON. Their inclusion would better establish the claimed competitiveness on image try-on. Since Table 2 also compares image try-on followed by animation, the authors should justify that baseline selection or extend the comparison using the same animation protocol. Such video comparisons would assess the complete image-plus-animation pipelines, so their interpretation should account for the animation stage.

  3. BiCorr-Test provides limited evidence for robustness, and direct evaluation on the source benchmarks is missing. BiCorr-Test contains 150 short clips selected from the official Eevee and ViT-HD test splits: 123 from Eevee and 27 from ViT-HD. It favors clear garment details and predominantly frontal views, explicitly excluding back and strict profile views. Why not evaluate the same trained model directly on the complete official Eevee and ViT-HD test splits and report each dataset separately? The added value of BiCorr-Test as a separate benchmark contribution also needs clearer justification beyond its curation from existing test data.

  4. The user study does not directly report how often participants prefer each method. Appendix A.3 provides randomized, anonymized five-point ratings and a mixed-effects analysis, which are useful evidence. However, the mean-score differences in Table 5 do not directly show how often users select BiCorr as the best output or prefer it to the strongest baseline. An explicit best-output choice or pairwise preference task, allowing ties, would provide a useful complementary assessment. At minimum, the authors could report within-trial win/tie/loss rates derived from the existing ratings, clearly distinguishing these from explicitly elicited preferences, together with confidence intervals and the number of cases and judgments per comparison. Since each trial includes BiCorr and only one to three baselines, winner rates should use matched candidate sets or account for the differing competitor sets. This would make the claimed perceptual advantage easier to assess.

  5. The additional value of LHFSC for garment-detail evaluation is not yet sufficiently established. LHFSC compares high-frequency Fourier magnitudes while discarding phase, so spectral agreement alone does not establish that lettering or motifs have the correct spatial structure. Appendix C.4 also acknowledges that the agnostic mask may include non-garment content and that high-frequency responses can arise from folds, boundaries, or noise. Table 6 shows only small improvements over LPIPS in within-case correlation (0.847 versus 0.839) and ranking agreement (89.3% versus 88.4%), with confidence intervals for the differences spanning zero. Controlled evaluations involving incorrect lettering, displaced motifs, blur, and added noise would clarify the metric's sensitivities and blind spots. Its intended role as a complementary measure is reasonable, but the present evidence does not establish a clear additional advantage for distinguishing garment-detail quality within the same test case.

Questions for authors

My questions are incorporated into the five points in Critical Weaknesses. Please respond to them individually. Clarifications concerning the ablation controls, evaluation on the original Eevee and ViT-HD test splits, and the additional value of LHFSC could materially affect my assessment.

11 comments
AuthorsOpenReview ↗

Response to Reviewer 8FQU (Image and video baseline comparisons)

We thank the reviewer for the constructive feedback and recognition of BiCorr's correspondence-learning framework and empirical improvements. We address the remaining concerns below with additional experiments and clarifications.

1. Controlled ablations, hard-target behavior, and attribution of generation improvements (W1)

The requested controlled ablations require additional training runs and evaluation. These experiments are currently underway, and we expect to report the results in a follow-up comment within three days, before the discussion period closes.

2. Comparison with recent image-based try-on methods and image-to-video baselines (W2)

We added Mobile-VTON and RefTon evaluations on VITON-HD and DressCode upper-body subset (DressCode-UB) under paired and unpaired settings. For RefTon, we use its inference mode without optional extra person-reference images to match the person–garment inputs available to BiCorr. Table W2-A shows that BiCorr remains competitive with methods specifically designed for image-based try-on, particularly in FID and KID, while Mobile-VTON achieves better SSIM and LPIPS.

Garments2Look introduces a multi-reference, outfit-level dataset and benchmark covering multiple garments, accessories, layering, and styling. Its task differs from our single-garment setting. We will discuss this distinction in Related Work in the revision.

Table W2-A. Image-level comparison. All images are evaluated at 1024 × 768. KID is multiplied by 10310^3. pp/uu denote paired/unpaired. Bold: best; underline: second best.

DatasetMethodSSIM!↑\mathrm{SSIM}\\!\uparrowLPIPS!↓\mathrm{LPIPS}\\!\downarrowFIDp!↓\mathrm{FID}^{p}\\!\downarrowKIDp!↓\mathrm{KID}^{p}\\!\downarrowFIDu!↓\mathrm{FID}^{u}\\!\downarrowKIDu!↓\mathrm{KID}^{u}\\!\downarrow
VITON-HDBiCorr0.8985‾\underline{0.8985}0.0540‾\underline{0.0540}5.31500.50208.8190‾\underline{8.8190}0.7560
Mobile-VTON0.90820.05035.59310.70789.72351.3626
RefTon0.89050.05895.4793‾\underline{5.4793}0.6117‾\underline{0.6117}8.66070.8362‾\underline{0.8362}
DressCode-UBBiCorr0.9497‾\underline{0.9497}0.0222‾\underline{0.0222}5.74160.265210.17970.5115
Mobile-VTON0.95570.02115.8215‾\underline{5.8215}0.2322‾\underline{0.2322}10.7071‾\underline{10.7071}0.7368‾\underline{0.7368}
RefTon0.93670.03276.55880.208711.07230.7670

For video comparisons, we originally selected CORAL as a correspondence-aware baseline and Any2AnyTryon as a general image-based try-on baseline. We extend the comparison with Mobile-VTON and RefTon on 50 shared upper-body cases, and evaluate RefTon on 150 cases across all three garment categories. Mobile-VTON is included only in the upper-body comparison, in accordance with its released upper-body evaluation protocol.

All image-based try-on methods use the same StableAnimator (SA) checkpoint and configuration (832 × 624 resolution, 64 frames, and 25 fps), whereas BiCorr generates videos directly without an additional animation stage. BiCorr achieves the best results on all seven reported video metrics in both comparisons. Table W2-B reports the three-category results; Tables 14–16 (Appendices D.2–D.3 of the revised manuscript) provide the upper-body comparison and reference-frame diagnostics. These results compare complete image-plus-animation pipelines with direct video generation. Differences therefore reflect both image try-on and animation, rather than image-model quality alone.

Table W2-B. Raw video metrics on 150 cases across three garment categories (50 per category). All image-based methods are followed by the same SA pipeline; BiCorr generates videos directly. pp/uu denote paired/unpaired. Bold: best; underline: second best.

Video SystemVFIDIp!↓\mathrm{VFID}_{I}^{p}\\!\downarrowVFIDRp!↓\mathrm{VFID}_{R}^{p}\\!\downarrowSSIMp!↑\mathrm{SSIM}^{p}\\!\uparrowLPIPSp!↓\mathrm{LPIPS}^{p}\\!\downarrowLHFSCp!↑\mathrm{LHFSC}^{p}\\!\uparrowVFIDIu!↓\mathrm{VFID}_{I}^{u}\\!\downarrowVFIDRu!↓\mathrm{VFID}_{R}^{u}\\!\downarrow
BiCorr9.70780.18080.88010.07150.745818.68190.3542
CORAL + SA14.7538‾\underline{14.7538}0.2083‾\underline{0.2083}0.8044‾\underline{0.8044}0.1178‾\underline{0.1178}0.619122.42790.4422‾\underline{0.4422}
Any2AnyTryon + SA16.69460.42180.79610.12940.6337‾\underline{0.6337}24.11600.6828
RefTon + SA15.09410.24000.79470.12890.606021.6908‾\underline{21.6908}0.4649
AuthorsOpenReview ↗

Response to Reviewer 8FQU (Benchmark scope and user study)

3. BiCorr-Test scope and evaluation on the original source benchmarks (W3)

We agree that BiCorr-Test alone cannot establish robustness across viewpoints. Our main video benchmark is ViViD-S. BiCorr-Test complements it with a reproducible, detail-oriented evaluation protocol designed to assess garment-detail fidelity and overall video quality across three garment categories. Clip identities and intervals were fixed before inspecting outputs or scores. We will clarify its contribution as a curated protocol for controlled evaluation, rather than a newly collected dataset or a comprehensive robustness benchmark.

Regarding the viewpoint selection, we follow the single frontal-reference setting: back-facing and strict-profile views may expose garment regions absent from the reference image, making fine-grained evaluation of details such as logos and patterns underconstrained. CatV2TON adopts a similar protocol, filtering back-facing frames and using 64 consecutive frontal frames for ViViD-S (Chong et al., 2025).

We are evaluating the same BiCorr checkpoint, without further fine-tuning, on all 1,000 official Eevee test IDs against MagicTryOn. Following the Eevee evaluation setting, we use 49-frame clips at 624 × 832 resolution. Evaluation is ongoing, with completion expected in approximately two days. We will report the Eevee results separately. Complete evaluation on the official ViT-HD test split exceeds the remaining rebuttal budget and is planned as follow-up work.

4. Rating-derived win/tie/loss analysis of the user study (W4)

We reanalyzed the existing anonymized five-point ratings to obtain within-trial win/tie/loss (W/T/L) rates against MagicTryOn, the strongest baseline in our user study. Each comparison matches the participant, trial, case, and criterion. Higher, equal, and lower BiCorr ratings count as wins, ties, and losses, respectively. These are rating-derived comparisons, not explicit preference votes.

Table W4. Rating-derived within-trial comparisons between BiCorr and MagicTryOn. W/T/L rates are computed over matched judgments, with ties included in the denominator. All intervals are pointwise 95% CIs; pp denotes percentage points.

ModeCriterionMatched Comparisons / Cases / ParticipantsW/T/L CountsWin % [95% CI]Tie % [95% CI]Loss % [95% CI]Net W−L, pp [95% CI]
PairedGarment detail119 / 35 / 6057/48/1447.9 [29.9, 65.2]40.3 [24.8, 57.3]11.8 [3.3, 23.8]36.1 [10.6, 58.8]
PairedOverall quality119 / 35 / 6060/44/1550.4 [32.3, 67.8]37.0 [21.5, 54.7]12.6 [3.6, 25.2]37.8 [12.2, 60.4]
PairedTemporal consistency119 / 35 / 6046/62/1138.7 [22.6, 56.0]52.1 [36.1, 68.3]9.2 [1.9, 19.6]29.4 [7.8, 50.8]
UnpairedGarment detail91 / 33 / 6045/32/1449.5 [30.1, 68.8]35.2 [18.9, 52.6]15.4 [4.4, 29.5]34.1 [4.8, 60.9]
UnpairedOverall quality91 / 33 / 6045/34/1249.5 [30.1, 68.4]37.4 [21.4, 53.9]13.2 [2.8, 26.8]36.3 [7.8, 62.4]
UnpairedTemporal consistency91 / 33 / 6033/49/936.3 [18.5, 55.7]53.8 [35.2, 71.7]9.9 [1.3, 22.4]26.4 [1.2, 49.5]

Intervals use 5,000 two-way participant–case cluster bootstrap replicates, independently resampling participants and cases while retaining observed within-trial comparisons. Net-rate intervals are computed directly from bootstrap differences. BiCorr wins outnumber losses across all criteria and both protocols.

Only trials presenting both methods enter each pairwise comparison. We will also report rating-derived unique-highest, tied-highest, and non-highest frequencies separately for each exact candidate set, without pooling across sets, alongside pairwise comparisons with the remaining baselines.

5. Sensitivity, limitations, and complementary value of LHFSC (W5)

The additional LHFSC sensitivity tests and human-judgment analyses are currently underway. We expect to report the results, together with confidence intervals, in a follow-up comment within three days, before the discussion period closes.

My main concern remains unresolved. Although this is a video virtual try-on (VVT) paper, its core methodological designs are largely adapted from existing image-based VTO approaches, without sufficient video-specific innovations to explicitly address temporal consistency and video dynamics.

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

This paper addresses the problem that fine-grained patterns and local structures in video virtual try-on (VVT) are prone to distortion and misplacement under large motion and occlusion. It proposes BiCorr, which explicitly learns bidirectional latent correspondence between person-video tokens and reference-garment tokens in a garment-conditioned video DiT. The authors construct offline pseudo-correspondences via DINOv3 direct matching combined with CoTracker3 trajectory propagation, map them onto the compressed latent token grid, and supervise attention using Gaussian-based distributional targets (rather than single-point one-hot targets): in the P→G direction, the targets guide person tokens to retrieve features from the correct garment regions, while in the reverse G→P direction they guide garment tokens to localize their spatiotemporal positions in the video. All matching/tracking/correspondence supervision is used only during training, with zero additional inference overhead. The authors further introduce a new metric, LHFSC (Local High-Frequency Spectral Consistency), and curate their own BiCorr-Test evaluation set.

Critical strengths

1.The problem is real and well-motivated. "Attention retrieves features from non-corresponding garment regions, causing detail distortion" is a concrete, practically valuable pain point in video virtual try-on. The motivation is clearly articulated rather than manufactured for the sake of publication. 2.There is a genuinely video-specific technical insight. Naively porting correspondence supervision from image to video DiTs hits the ambiguity that temporal compression assigns multiple garment locations to the same token. Handling this with Gaussian-mixture distributional targets that retain distinct garment locations as separate components is the paper's most valuable, non-trivial contribution — not a trivial transplant. 3.Experiments are solid and comprehensive. The evaluation spans multiple benchmarks (ViViD-S / BiCorr-Test / VITON-HD / DressCode), paired and unpaired protocols, quantitative and qualitative results, a 60-participant user study, and three-dimensional ablations (supervision direction / target formulation / pseudo-correspondence construction). The evidence chain is fairly complete.

Critical weaknesses
  1. Effectiveness heavily depends on pseudo-correspondence quality. The entire supervision stack is built on offline pseudo-labels from DINOv3 matching + CoTracker3 tracking. Noise in these pseudo-labels directly contaminates supervision — a limitation the authors themselves acknowledge. There is no sensitivity analysis linking pseudo-correspondence quality to final performance, which is a gap.
  2. Benchmark carries self-construction risk. BiCorr-Test is curated by the authors from official splits and dominated by short, frontal-view clips, introducing selection bias; LHFSC is a self-proposed metric (measures magnitude only, not phase) with limited persuasiveness despite the reported ρ = 0.719.
  3. Weak theoretical contribution. The method is empirical, offering no deep theoretical insight or transferable principled conclusion.
Questions for authors

See weaknesses

4 comments
Reviewer_BQxpOpenReview ↗

I mixed up this paper with previous one

I'm sorry, i recieved 5 papers and i mixed up this one with the previous, please ignore my review

AuthorsOpenReview ↗

Response to Reviewer BQxp (Pseudo-correspondence quality and benchmark evaluation)

We thank the reviewer for the constructive feedback and recognition of BiCorr's video-specific insight and comprehensive evaluation. We address the three concerns below with additional experiments and clarifications.

1. Sensitivity to pseudo-correspondence quality (W1)

We have designed controlled pseudo-correspondence sensitivity experiments, and training and inference are underway. We aim to provide the results and analysis in a follow-up comment before the discussion deadline.

2. Benchmark selection and LHFSC validity (W2)

We conducted additional evaluations to examine whether our findings generalize beyond the selected test subset and whether LHFSC provides complementary information to established metrics.

BiCorr-Test and selection bias.

BiCorr-Test complements our primary benchmark, ViViD-S, with a detail-oriented evaluation of garment fidelity and overall video quality across three garment categories. We favor frontal views because back-facing and strict profile views may expose garment regions absent from the reference image, making fine-grained evaluation of logos and patterns unreliable. CatV2TON [1] adopts a similar protocol, filtering back-facing frames and selecting 64 consecutive frontal frames for ViViD-S. Importantly, clip identities and intervals were fixed before inspecting model outputs or scores, avoiding result-driven selection. To assess whether our improvements depend on this curation, we evaluated the same BiCorr checkpoint on the complete official Eevee test split (1,000 IDs), without further fine-tuning.

Table W2-A. Quantitative comparison on the complete official Eevee test split. Baseline results are taken from Table 2 of the original Eevee paper [2] at the corresponding resolution. Superscripts p/u denote paired/unpaired settings. Bold: best; underline: second best. † indicates fine-tuning on the Eevee training split and the use of additional garment-detail reference images.

MethodVFIDIp!↓\mathrm{VFID}_{I}^{p}\\!\downarrowVFIDRp!↓\mathrm{VFID}_{R}^{p}\\!\downarrowSSIMp!↑\mathrm{SSIM}^{p}\\!\uparrowLPIPSp!↓\mathrm{LPIPS}^{p}\\!\downarrowVFIDIu!↓\mathrm{VFID}_{I}^{u}\\!\downarrowVFIDRu!↓\mathrm{VFID}_{R}^{u}\\!\downarrow
ViViD9.2110.2850.8710.11112.1940.389
MagicTryOn6.167‾\underline{6.167}0.1060.908‾\underline{0.908}0.0749.8650.161
Eevee†6.5200.0760.9050.071‾\underline{0.071}9.777‾\underline{9.777}0.219
BiCorr5.5150.089‾\underline{0.089}0.9130.0639.5740.170‾\underline{0.170}

As shown in Table W2-A, BiCorr outperforms MagicTryOn on all four paired metrics and achieves the best unpaired VFID-I. It also surpasses the Eevee model, which uses dataset-specific fine-tuning and additional garment references, on three of four paired metrics and both unpaired metrics. Together with our ViViD-S results, this evaluation demonstrates that BiCorr's improvements are not confined to the curated subset.

LHFSC and metric validity.

Magnitude-only design and phase sensitivity. Video virtual try-on involves substantial motion and non-rigid garment deformation, which can introduce small texture displacements. LHFSC therefore intentionally compares local high-frequency Fourier magnitudes while omitting phase, aiming to capture texture degradation without over-penalizing such displacements. To validate this design, we compared LHFSC with a phase-aware counterpart under identical preprocessing and normalization (Appendix D.5, Fig. 13 and Table 17). For a 1-pixel translation, LHFSC retains a similarity of 0.9962, whereas the phase-aware score drops to 0.5534. A smooth warp with a maximum displacement of 1 pixel yields a similar contrast (0.9865 vs. 0.7585). These results illustrate the benefit of omitting phase for tolerating small motion-induced texture displacements.

Complementary value and human validation. LHFSC provides a complementary measure of local high-frequency texture fidelity rather than replacing established perceptual metrics. To examine its distinct sensitivity, we conducted controlled tests involving blur, noise, and translation across 12 garment cases, along with additional local motif and lettering edits (Appendix D.4, Fig. 12). As shown in Fig. 12(b), LHFSC is more sensitive to blur than LPIPS and SSIM, while exhibiting a different response to noise, highlighting its complementary sensitivity to high-frequency texture degradation.

References.

[1] Chong et al. (2025). CatV2TON: Taming Diffusion Transformers for Vision-Based Virtual Try-On with Temporal Concatenation. arXiv:2501.11325.

[2] Zeng et al. (2026). Eevee: Towards Close-up High-resolution Video-based Virtual Try-on. CVPR 2026, pp. 4614–4624.

AuthorsOpenReview ↗

Response to Reviewer BQxp (LHFSC validation and theoretical insights)

We continue our response to Weakness 2 below.

2. Benchmark selection and LHFSC validity (W2)

More importantly, we evaluated whether LHFSC provides incremental predictive value beyond established metrics. We selected regularization by leave-one-case-out validation on 35 calibration cases, then refitted and froze the predictors before collecting new judgments. Twenty assessors provided blinded pairwise judgments on 295 disjoint test cases, yielding 1,199 judgments, including 1,087 non-tied preferences. As shown in Table W2-B, adding LHFSC improves held-out agreement with human garment-detail preferences from 89.60% to 91.35% over LPIPS, and from 88.97% to 91.35% over LPIPS + SSIM. Both paired video-bootstrap 95% CIs exclude zero, providing evidence of LHFSC's incremental predictive value beyond the originally reported correlation (ρ = 0.719).

Table W2-B. Agreement with held-out human garment-detail preferences. Agreement measures predicted-order agreement with non-tied human preferences, averaged across assessors, method pairs, and videos hierarchically. Of 295 cases, 294 contribute non-tied judgments. Paired video-bootstrap 95% CIs use 20,000 dataset- and category-stratified replicates with the assessor panel fixed. pp denotes percentage points.

Baseline predictor featuresAgreement ↑With LHFSC ↑Δ agreement (pp)95% CI
Full-frame LPIPS89.60%91.35%+1.75[+0.23, +3.30]
Full-frame LPIPS + SSIM88.97%91.35%+2.38[+0.68, +4.07]

Together, these evaluations support BiCorr-Test as a targeted complement to standard benchmarks and LHFSC as a complementary garment-detail metric. We will incorporate the additional analyses and clarify the scope of both evaluation components in the revised version.

We next address the methodological insights.

3. Theoretical insights and potential generalizability (W3)

We agree that BiCorr is primarily a problem-driven contribution, motivated by improving garment-detail fidelity and temporal consistency in video virtual try-ons. Nevertheless, our work provides insights into correspondence supervision for temporally compressed video models.

Correspondence learning under temporal compression. Temporal compression presents a general challenge for fine-grained correspondence learning. For example, DiffTrack [3] avoids temporal compression (q=1) to improve point-tracking precision, while MVTrack4Gen [4] supervises individual correspondences within each target frame. Although these works address different settings, they highlight the importance of adapting correspondence learning to the underlying temporal representation.

In BiCorr, observations from different frames may map to the same latent token while corresponding to distinct garment regions. This ambiguity occurs in practice: 20.51% of valid mapped person-video queries in ViViD-S have multiple garment correspondences before query selection and top-K truncation (Appendix A.4 of the main paper). Rather than removing temporal compression or selecting a single correspondence, we retain multiple supported locations through Gaussian-mixture targets. Our ablation (Table 3 of the main paper) shows that the complete distributional-target formulation outperforms hard single-point targets on LPIPS, LHFSC, and all four VFID metrics. Together, these observations suggest that correspondence supervision should account for the model's latent tokenization and preserve valid alternatives when multiple observations share a compressed representation.

BiCorr further supervises G→P attention over the full video-token domain, jointly guiding cross-frame allocation and within-frame localization. Our ablations (Table 3) show that adding G→P supervision improves both correspondence accuracy and generation quality, supporting the practical value of this formulation.

Potential generalizability. These insights concern correspondence mapping and attention supervision rather than Wan2.1-specific operations [5]. They may inform other reference-guided video generation tasks, such as video editing and animation, where compressed video tokens must retrieve corresponding reference information across motion. However, cross-backbone and cross-task transfer has not yet been experimentally validated and remains future work. We will clarify these insights and their potential applicability in the revised version.

References.

[3] Nam et al. (2025). Emergent Temporal Correspondences from Video Diffusion Transformers. arXiv:2506.17220.

[4] Lee et al. (2026). MVTrack4Gen: Multi-View Point Tracking as Geometric Supervision for 4D Video Generation. arXiv:2606.26087.

[5] Team Wan et al. (2025). Wan: Open and Advanced Large-Scale Video Generative Models. arXiv:2503.20314.

AuthorsOpenReview ↗

Response to Reviewer BQxp (Pseudo-correspondence quality and sensitivity analysis)

We provide the following updates to our previous responses to Points 1.

1. Sensitivity to pseudo-correspondence quality (W1)

We agree that pseudo-correspondence quality is important for effective attention supervision.

Reliability and quality assessment. During pseudo-correspondence construction, BiCorr applies reliability filtering and validation to direct and tracked matches, together with coverage-balanced query selection to avoid concentrating supervision on easily matched frames (Sec. 3.2). To quantitatively assess correspondence quality at different stages, we conducted an independent, method-blinded human audit of 3,244 unique retained correspondences from 470 frames across 48 video clips. The table includes 3,358 method-specific assignments because 114 correspondences are shared across methods. An independent assessor, blinded to method identities and confidence scores, labeled each correspondence as correct or incorrect. The audit provides quantitative evidence of retained-match reliability, with Fusion achieving 84.98% video-equal precision.

Table W1-A. Independent correspondence audit. The number of videos indicates clips containing audited matches for each method. Video-equal precision averages correctness across queries, frames, and videos, giving each video equal weight. The 95% CIs use video-cluster bootstrap resampling.

MethodVideos with Audited MatchesCorrect / IncorrectVideo-Equal Precision (%) ↑95% CI (%)
Match481,104 / 59074.41[68.85, 79.82]
Track41606 / 9587.04[83.85, 90.10]
Fusion44813 / 15084.98[81.67, 87.94]

Sensitivity analysis. However, the audit does not establish how pseudo-correspondence quality affects generation performance. To address this, we designed controlled experiments that either lower the thresholds for matching and filtering to admit more low-confidence correspondences or randomly perturb matched person–garment pairs, followed by training and evaluation under matched settings.

These experiments require approximately 1 day for offline pseudo-correspondence extraction, 5 days for full three-stage training of each 1.3B variant on 4 NVIDIA H20 GPUs, and 6 hours for inference on 180 paired and 180 unpaired test videos. Given these time and computational requirements, we could not complete the experiments within the discussion period. We will include this sensitivity analysis in the revised paper.

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

The paper presents BiCorr, a bidirectional latent correspondence learning framework for video virtual try-on. It uses DINOv3 and CoTracker3 to construct high-confidence pseudo-correspondences offline, which supervise attention distributions for both person-to-garment retrieval and garment-to-person spatiotemporal localization. Gaussian-based distributional targets accommodate local positional uncertainty and retain multiple garment matches associated with the same temporally compressed person-video token. The method improves garment-detail fidelity and achieves leading performance on the reported public video benchmarks without additional inference overhead. The paper also introduces BiCorr-Test and local high-frequency spectral consistency (LHFSC) as complementary resources for evaluating fine-grained garment fidelity.

Critical strengths
  1. The paper presents a coherent correspondence-learning framework for improving local garment fidelity without additional inference overhead, making the approach practically useful.

  2. Evaluations across public benchmarks, BiCorr-Test, ablations, and a user study provide complementary evidence of improved garment detail and overall quality, strengthening confidence in the method’s effectiveness.

Critical weaknesses
  1. More targeted experiments are needed to justify the complex loss formulation, quantify the accuracy and construction cost of the pseudo-correspondences, and assess performance on newer video generation backbones and more challenging video try-on scenarios. In particular, it remains unclear whether the method provides meaningful gains over stronger backbones, or whether pseudo-correspondence accuracy degrades under camera motion and large changes in body pose. These gaps limit confidence in the method’s practical cost, robustness, and broader applicability.

2.The introduction suggests that retrieving features from incorrect garment regions causes detail distortion, but the paper provides limited evidence that this is a key factor limiting garment fidelity. Figure 4 offers useful qualitative support, yet hard single-point targets achieve higher localization and coverage scores while producing worse results. This leaves unclear which aspects of attention alignment improve garment details and how much of the gain comes from better feature retrieval versus other effects of the auxiliary supervision.

Questions for authors

1.Why are the additional loss terms needed beyond cross-entropy? The current ablations do not show how much each component contributes. A comparison with a cross-entropy-only baseline, followed by adding the main components individually, would help explain which parts of the design drive the gains. 2.How reliable are the pseudo-correspondences produced by direct matching, tracking, and fusion? Could you show some successful and failed matches, and quantify their accuracy on a small, independently checked subset? It would also help to know how long the offline construction takes and what hardware it requires. 3. What evidence supports attention misalignment as a key cause of garment-detail distortion? In particular, why do hard single-point targets achieve better localization and coverage scores but poorer generation quality? Could a controlled analysis during generation, using consistent attention readout settings across variants, show whether correcting misplaced attention improves the corresponding garment details? 4.How dependent is BiCorr on the Wan2.1 architecture? Could you clarify which parts would need to change when adapting it to another backbone? Testing whether the gains carry over to models such as Wan2.2 or MiniMax H3 would be useful future work.

5 comments
AuthorsOpenReview ↗

Response to Reviewer MPUy (Correspondence reliability and attention analysis)

Thank you for recognizing BiCorr's correspondence-learning framework and its practical benefits. We address your questions below with additional experiments and clarifications.

1. Necessity and individual contributions of the additional loss terms

The requested CE-only and component-wise ablations require additional training runs and evaluation. These experiments are currently underway, and we expect to provide the results in a follow-up comment within the next few days, before the discussion period closes.

2. Reliability, failure cases, and computational cost of pseudo-correspondence construction

We independently inspected the pseudo-correspondences, added successful and failed examples, and measured the offline construction cost.

We evaluated three stages of pseudo-correspondence construction: direct frame-wise DINOv3 matches (Match), matches propagated from DINOv3 seeds to non-anchor frames using CoTracker3 (Track), and matches retained after quality-gated fusion (Fusion).

The evaluation covers 3,244 unique retained correspondences from 470 frames across 48 person–garment clips, balanced across three garment categories. An independent assessor, uninvolved in correspondence extraction or parameter tuning, labeled each pair as correct or incorrect. Method identities and confidence scores were hidden.

Table Q2-A. Independent correspondence audit.

MethodVideos with Audited MatchesCorrect / IncorrectVideo-Equal Precision (%) ↑95% CI (%)
Match481,104 / 59074.41[68.85, 79.82]
Track41606 / 9587.04[83.85, 90.10]
Fusion44813 / 15084.98[81.67, 87.94]

Match provides the broadest frame coverage, whereas Track has the highest precision. Fusion retains broader coverage than Track and higher precision than Match, providing a trade-off between reliability and coverage. Precision is computed hierarchically by averaging across queries, frames, and videos, with equal weight assigned to each video. It therefore differs from the pooled count ratio. The 95% confidence intervals (CI) use 5,000 source/category-stratified video-cluster bootstrap replicates. These results measure retained-match precision rather than recall. The table contains 3,358 method-specific assignments because 114 pairs are shared across methods.

Figure 11 (Appendix D.1) shows 16 audited correspondences from four frames, illustrating successful garment-detail matches, a residual motif mismatch in Fusion, a Track error inherited from an incorrect DINOv3 seed, and Match-stage sleeve-to-bodice mismatches not retained by Track or Fusion.

Given preprocessed videos and visible-garment masks, correspondence construction takes 62.12 s per 64-frame clip on average (range: 23.42–111.02 s) on a single NVIDIA RTX 4090 (24 GB VRAM), with a sampled peak GPU-process memory of 6.82 GiB. Timing includes model initialization, matching, tracking, quality scoring, fusion, refinement, and output validation, but excludes upstream mask and DensePose preprocessing. Correspondences are constructed offline and cached, introducing no additional inference-time overhead.

This evaluation quantifies precision among retained correspondences but does not establish robustness under severe camera motion or large pose changes. We will acknowledge this limitation and explicitly identify systematic evaluation under these challenging conditions as future work in the revised manuscript.

3. Attention alignment and its relationship to garment-detail fidelity

We agree that our explanation of why Hard (hard single-point targets) achieves better localization and coverage but poorer generation quality may have led to a misleading causal interpretation. We do not claim that attention misalignment is the key cause of garment-detail distortion; rather, it is one potential contributing factor. To this end, we conducted a controlled generation analysis and an attention intervention.

Attention analysis of the Full–Hard ablation. We compared Full (the complete BiCorr method) and Hard on 173 ViViD-S videos with valid pseudo-correspondences. Both used identical initial latents, sampling settings, queries, and evaluation pseudo-labels. For both variants, post-softmax attention was read out at denoising steps 0/6/12/18/24 and averaged across heads and all 30 blocks, with equal weight assigned to each block.

AuthorsOpenReview ↗

Response to Reviewer MPUy (Attention intervention and backbone generalization)

3. Attention alignment and its relationship to garment-detail fidelity (continued)

Table Q3-A. Common-readout comparison. Differences are Hard−Full, where the results in brackets show paired-video bootstrap 95% CIs.

MeasureFullHardHard−Full (95% CI)
Self-P→G PCK@2 (%) ↑52.6854.58+1.90 [1.14, 2.66]
Self-P→G mode coverage (MC, %) ↑19.7221.05+1.33 [1.08, 1.57]
Effective location count ↑1.91621.8676−0.0486 [−0.0524, −0.0448]
LHFSC ↑0.668490.65336‾\underline{0.65336}−0.01513 [−0.02105, −0.00921]
LPIPS ↓0.084440.08679‾\underline{0.08679}+0.00235 [0.00169, 0.00306]

We additionally measure the entropy-based effective location count. For each query, we normalize the attention mass across valid correspondence regions and compute the exponential of its entropy. This measures the effective number of attended correspondence locations: it approaches 1 when attention concentrates on a single location and increases when attention is distributed more evenly across multiple valid locations. Unlike PCK@2, which measures peak localization, or MC, which measures capped coverage, this metric specifically captures the relative balance of attention across valid correspondences.

Although Hard achieves higher PCK@2 and MC, Full has a higher effective location count (1.9162 vs. 1.8676), consistently across 172/173 videos. This indicates more balanced attention across alternative garment correspondences, which is particularly relevant when temporal compression maps multiple valid garment locations to the same person-video token. This difference in attention allocation may help explain Full's better garment-detail fidelity, motivating the attention intervention below.

Generation-time intervention. We transferred Full-derived attention weights into Hard on 180 paired videos (60 per garment category). At each denoising step, Full computes attention weights from Hard's current latent. These weights replaced Hard’s original weights in all self- and conditioning cross-attention branches across 30 blocks and 25 steps. Hard's value projections and other model parameters remain unchanged, and both runs use identical initial latents.

Table Q3-B. Attention intervention. Differences are relative to Hard. Brackets show paired-video bootstrap 95% CIs.

MeasureHard BaselineHard + Full AttentionDifference [95% CI]Improved Videos
LPIPS ↓0.08626‾\underline{0.08626}0.08459−0.00167 [−0.00218, −0.00121]140/180
LHFSC ↑0.65440‾\underline{0.65440}0.66126+0.00686 [+0.00312, +0.01048]124/180
SSIM ↑0.86538‾\underline{0.86538}0.86580+0.00042 [−0.00001, +0.00084]115/180
VFIDI!↓\mathrm{VFID}_{I}\\!\downarrow12.1950‾\underline{12.1950}11.6770−0.5180—
VFIDR!↓\mathrm{VFID}_{R}\\!\downarrow0.23970.2654‾\underline{0.2654}+0.0257—

Both LPIPS and LHFSC improve significantly, with their paired-video bootstrap 95% confidence intervals excluding zero. SSIM shows a favorable improvement, while VFID results are mixed. The intervention provides further evidence that learned attention weights contribute to reconstruction quality. We will clarify the discussion to distinguish the proposed explanation of the Hard–Full gap from a demonstrated causal mechanism.

4. Backbone independence, required adaptations, and cross-architecture generalization

BiCorr is architecture-agnostic by design. Its core operations are independent of the specific backbone architecture rather than being tailored to or relying on Wan2.1. As illustrated by the three lower panels of Figure 2 in our paper, pseudo-correspondence construction operates directly on images and frames and is independent of the backbone. Correspondence target construction represents matched locations as distributions over tokens. Bidirectional correspondence supervision aligns the resulting attention distributions with these targets. Neither mechanism relies on Wan2.1-specific operations.

Adapting BiCorr to another backbone therefore mainly requires modifying two implementation-level components. The first is mapping frame-level correspondences to the backbone's token grids while accounting for its VAE compression and patchification. The second is extracting person–garment attention according to its token layout, normalization, positional encoding, and attention masks. The distributional target formulation and bidirectional supervision objectives remain unchanged.

We have not evaluated Wan2.2 or MiniMax H3 in this submission because implementing these adaptations, fine-tuning, and conducting matched baseline comparisons exceeded the available rebuttal time and compute budget. Nevertheless, given the architecture-independent formulation of BiCorr, we expect its core mechanisms to transfer to other backbones with appropriate implementation-level adaptations. Systematically validating this transferability on Wan2.2 and MiniMax H3 is an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer MPUy (Additional loss ablations)

Thank you again for your constructive feedback. We provide the following update to our previous response to Point 1.

1. Necessity and individual contributions of the additional loss terms (Additional experimental results)

Cross-entropy (CE) provides distributional correspondence supervision, while three auxiliary terms offer complementary constraints: gated entropy encourages concentrated P→G attention with reduced strength for uncertain targets, shape KL aligns relative G→P attention probabilities within the correspondence support, and outside-region penalties suppress attention leakage beyond the support. We conducted additional complete-training and component-wise experiments to examine their contributions.

CE-only comparison. We trained CE-only using the same curriculum and settings as Full, evaluating both on ViViD-S (180 paired and 180 unpaired videos).

Table Q1-A. CE-only versus Full after complete three-stage training. Subscripts p/u denote paired/unpaired evaluation.

VariantVFIDIp!↓\mathrm{VFID}_{I}^{p}\\!\downarrowVFIDRp!↓\mathrm{VFID}_{R}^{p}\\!\downarrowSSIM!↑\mathrm{SSIM}\\!\uparrowLPIPS!↓\mathrm{LPIPS}\\!\downarrowLHFSC!↑\mathrm{LHFSC}\\!\uparrowVFIDIu!↓\mathrm{VFID}_{I}^{u}\\!\downarrowVFIDRu!↓\mathrm{VFID}_{R}^{u}\\!\downarrow
CE only11.91340.16150.86220.08510.666717.82190.2956
Full11.44430.1680.86310.08340.673317.74790.2846

Full outperforms CE-only on six of seven metrics, particularly LPIPS and LHFSC, confirming the overall benefit of our auxiliary losses.

Component-wise ablations. To examine individual contributions efficiently, we initialized all variants from the same generation-only (w/o corr) Stage-2 checkpoint and trained for 1,200 steps with the same data, settings, and correspondence targets, adding each auxiliary term separately to CE.

Table Q1-B. Matched component-wise ablations on ViViD-S. Subscripts p/u denote paired/unpaired evaluation.

VariantVFIDIp!↓\mathrm{VFID}_{I}^{p}\\!\downarrowVFIDRp!↓\mathrm{VFID}_{R}^{p}\\!\downarrowSSIM!↑\mathrm{SSIM}\\!\uparrowLPIPS!↓\mathrm{LPIPS}\\!\downarrowLHFSC!↑\mathrm{LHFSC}\\!\uparrowVFIDIu!↓\mathrm{VFID}_{I}^{u}\\!\downarrowVFIDRu!↓\mathrm{VFID}_{R}^{u}\\!\downarrow
CE only13.22550.87040.86570.08850.641118.23541.0126
+ Gated entropy13.52190.69880.86550.08840.64518.23040.7079
+ Shape KL13.08510.61350.86610.08810.643318.40920.6524
+ Outside-region penalties13.10210.60170.86570.08850.638718.21120.6279

Compared with CE-only, gated entropy improves 5 metrics, shape KL improves 6 metrics, and outside-region penalties improve 4 metrics. The remaining metrics are broadly comparable or show modest trade-offs.

Together, the complete-training results establish the overall benefit of combining auxiliary losses beyond CE, while the matched short-run ablations demonstrate their positive and complementary contributions. We will incorporate these findings into the revised version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.