acceptodds
Under review as a conference paper at ICLR 2027

W-DiT: Structure-Guided High-Frequency Propagation for Fashion Video Generation

Abstract

Generating fashion videos from reference images requires garment details to remain consistent under changing poses and cloth deformation. High-frequency product details, such as logos and fabric textures, should not be repeatedly regenerated across frames, as their appearance can progressively drift during motion. Our key insight is to generate low-frequency dynamics while propagating high-frequency product information directly from reference images along the garment's motion trajectories. Guided by this principle, we propose W-DiT, a video diffusion framework that couples generated garment motion with structure-guided detail propagation. First, frequency disentanglement separates low-frequency structure from high-frequency appearance and encodes reference details through a dedicated pathway that bypasses VAE compression. Second, structure-guided alignment matches low-frequency reference and target features to establish soft spatial correspondences that track the garment's evolving pose and deformation. Third, high-frequency propagation uses these correspondences to sample details directly from the fixed reference feature bank and inject them into each target frame, transporting product appearance along the motion trajectories. To support this task, we curate FashionVideo from authorized Xiaohongshu images, comprising approximately 10,000 synthesized fashion videos and 30,000 associated images with textual conditioning. Our experiments show that W-DiT preserves fine-grained garment details across changing poses while maintaining motion quality.

Primary area:
generative models
Keywords:
Image-to-Video generation, diffusion model

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_c6iMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes W-DiT, a reference-guided fashion video generation model built on top of Wan2.2. To keep garment textures and logos from drifting across frames, it decomposes reference images into low- and high-frequency components using a 2D Haar wavelet transform, uses the low-frequency structure to find spatial correspondences across video frames, and explicitly injects the high-frequency features into the DiT to boost texture accuracy. The authors also curate FashionVideo, a synthetic dataset of 10,000 videos, to train and evaluate the model alongside tests on Fashion-Text2Video.

Critical strengths

Decoupling low-frequency structural alignment from high-frequency texture injection is a reasonable and intuitive approach to preserving garment details. The experimental validation is fairly comprehensive, including a clean controlled ablation against Wan2.2 in Table 3 and zero-shot evaluation on real videos in Fashion-Text2Video. The visual results also show clear improvements in keeping logos and patterns consistent compared to the baselines.

Critical weaknesses
  1. My major concern is that this paper is essentially engineering a complex frequency patch onto an older backbone (Wan2.2) to fix a problem that SOTA video models (like Veo, Kling, or Omini) largely solve natively through scale, better VAEs, and full-attention conditioning. In fact, the authors' own dataset pipeline proves this point: while they claim existing models fail to keep garment details consistent, Section 4.2 (Lines 283–285) reveals that all 10k GT target videos in FashionVideo are synthesized by Wan2.5 from static images and used to train Wan2.2. If wan2.5 already preserves details well enough to serve as ground truth, then modern video models already solve this task without wavelet branches; if not, then the training target is fundamentally flawed. For this reason, I'm afraid this paper adds little value to the research community.
  2. The key evaluation metric Detail Fidelity Score (DFS) is used throughout the paper to prove that the paper's method can better preserve fine texture details and logos. However, DFS is just a DINOv2 cosine similarity score, which measures coarse semantic/appearance similarity. The paper has not made it clear why this is a good metric to evaluate high frequency textures.
Questions for authors
  1. Can you compare W-DiT against Wan2.5 (which generated the training targets) and modern SOTA video models like Veo or Omini-style unified DiTs on the test sets?
  2. For the try-on video editing application, it seems that the pipeline only edits the initial image and re-generates the whole video sequence. How does this method preserve the motion sequence and the camera framing/trajectory of the original video? If all the input is an edited image, then calling it "video editing" is misleading. Please clarify.
1 comment
AuthorsOpenReview ↗

Response to reviewer c6iM

[W1, Q1: Modern video models and training-data construction] FashionVideo uses human-screened Wan2.5 outputs. Candidates are retained only when they preserve garment details and facial identity and exhibit reasonable motion rather than remaining static. Only approximately 30% of Wan2.5-generated videos are usable under these criteria and pass dataset screening. Under the same criteria, W-DiT improves usability from 31% to 72% on 100 held-out reference-image and prompt pairs (Appendix B.1).

We extend the usability evaluation to Wan2.5, Veo 3, Omni, and Seedance 2.5, through their APIs on the same 100 held-out reference images and prompts and the joint criteria above:

MethodUsable / totalUsability rate
Wan2.531/10031.0%
Veo 337/10037.0%
Omni40/10040.0%
Seedance 2.545/10045.0%
W-DiT72/10072.0%

W-DiT exceeds the highest-scoring tested general-purpose model, Seedance 2.5, by 27 percentage points and Wan2.5 by 41 percentage points. These results show that preserving garment details and identity while producing reasonable motion remains challenging for the tested modern generators. The human-screened Wan2.5 targets supply successful training examples, and W-DiT substantially improves the reliability of meeting these requirements on held-out inputs. The main FashionVideo comparison also includes Kling (Table 1).

[W2: Detail evaluation beyond DFS] We complement garment-level DFS with human assessments and motion-compensated measurements of local details.

Human evaluation. Appendix B.2 evaluates whether local logo/pattern shape is preserved, placement remains anchored to the garment, and details remain stable during movement. W-DiT vs. Wan2.2 achieves pass rates of 78% vs. 38%, 85% vs. 52%, and 76% vs. 34%, respectively. Overall pass rates are 73/100 vs. 29/100. A separate pairwise study measures perceived garment-texture fidelity. W-DiT is preferred in 76.5% of non-tied judgments (Appendix A.2). Across 200 video pairs, the Pearson correlation between DFS and human texture preferences is r = 0.74, indicating agreement between the automated metric and human judgments (Appendix A.3).

Motion-compensated evaluation. Appendix D.2 evaluates detail consistency during motion using two complementary metrics. Garment Ewarp ↓ measures mean squared RGB differences between adjacent frames after optical-flow alignment, restricted to valid, non-occluded garment regions. Trajectory-Aligned DFS ↑ measures DINOv2 feature cosine similarity between tracked garment points and their corresponding reference-image points, assessing local reference fidelity along visible trajectories. With one reference, W-DiT vs. Wan2.2 obtains 3.41 vs. 5.18 (×10⁻³) for Garment Ewarp and 81.6 vs. 70.3 for Trajectory-Aligned DFS.

[Q2: Try-on video editing inputs] W-DiT directly accepts the original reference image and textual edit instructions, without requiring a pre-edited image or an external image-editing stage. At evaluation, Qwen-Image-Edit is used to prepare target-outfit inputs for comparison with general-purpose video generators. Appendix C.1 clarifies this evaluation setting, and Appendix C.2 reports both W-DiT's original-image results and the additional shared-input control. The generated motion and camera behavior follow the prompt. Frame-by-frame preservation of a source video's motion or camera trajectory is outside this experimental setting. Taking your suggestion into account, we have renamed the application garment-edited video generation to clarify its scope.

Reviewer_5VYXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes W-DiT, a structure-guided video diffusion framework for preserving fine-grained garment details during fashion video generation. The method separates low-frequency structure from high-frequency appearance using wavelet decomposition, establishes spatial correspondences through structure-guided attention, and propagates reference details into generated frames.

Critical strengths

1.The paper provides extensive comparisons with strong baselines, including Wan2.2, along with detailed ablation studies. 2.The separation of low-frequency motion and high-frequency appearance is intuitive and well-motivated. The proposed framework effectively integrates these components.

Critical weaknesses

1.The method mainly combines existing techniques, including wavelet decomposition, cross-attention, and gated feature injection, with limited methodological novelty.

Questions for authors

The structure-guided alignment relies on attention between reference and noisy target features. Could the authors provide quantitative evidence that these correspondences accurately track garment deformation across denoising steps?

5 comments
AuthorsOpenReview ↗

Response to reviewer 5VYX

[W1: Novelty] Existing fashion-video research has largely focused on two settings: (1) pose-driven animation that requires external pose sequences (e.g., DreamPose, ProFashion), and (2) video-conditioned virtual try-on that relies on a driving video (e.g., KeyTailor). In contrast, W-DiT targets a less-explored but practically important problem: generating motion directly from reference images and text, while preserving fine-grained product details (logos, patterns, fabric textures) under free pose change and cloth deformation-without any external pose or driving video.

At the methodological level, prior fashion-video methods have not, to our knowledge, investigated frequency decomposition as a mechanism for coordinating generated garment motion with persistent propagation of reference details. While wavelets and cross-attention appear in image restoration (e.g., HDW-SR) and gated feature injection is a known operation, the novelty of W-DiT lies in the frequency-specific formulation of the consistency problem itself: low-frequency structure is used to establish spatial correspondences that track the evolving garment, and high-frequency details are retrieved from a fixed reference bank and injected along those correspondences. This design explicitly separates "where to place details" from "what the details are," enabling detail reuse rather than repeated regeneration across frames.

[Q1: Quantitative evidence for correspondence accuracy] This is an insightful suggestion. To assess whether the learned correspondences accurately track garment deformation across denoising steps, we evaluate video-to-reference matching. For each of 100 videos, we manually mark the same five local garment landmarks (e.g., a logo corner) in the reference image and one clear generated frame, as in the point-level visualizations in Appendix B.7. Following DiffTrack (NeurIPS 2025), we use CoTracker to track the annotated garment points across video frames. At each sampled denoising step, we measure the Euclidean distance between the reference-attention peak for each visible tracked point and its annotated reference location. We normalize this error by the reference garment-box diagonal, obtained using Grounded-SAM under the DFS protocol. [email protected] reports the percentage of visible matches with normalized error at most 0.1. Scores are computed per video and averaged within each denoising stage (Appendix B.6).

MethodEarly: τ∈[0.7,1.0]Middle: τ∈(0.3,0.7)Late: τ∈[0.0,0.3]
HF + Standard Attention62.4%74.2%78.1%
W-DiT71.8%81.5%84.9%

The gains are 9.4, 7.3, and 6.8 percentage points, with the largest improvement in the high-noise interval. Appendix B.6 details the quantitative protocol, and Appendix B.7 provides point-level reference-attention visualizations.

Thank you for the detailed clarification. I agree that preserving fine-grained garment details without external pose sequences or driving videos is a meaningful problem, and I acknowledge the effectiveness of the proposed framework demonstrated by the experimental results.

However, I am still not fully convinced by the methodological novelty. The framework mainly combines existing techniques, including wavelet decomposition, cross-attention, and gated feature injection. Similar frequency-guided attention mechanisms have also been explored in previous diffusion-based methods, such as HDW-SR. While the authors emphasize the differences in the application setting and the integration of these components, I do not find a sufficiently clear methodological contribution beyond adapting existing techniques to fashion video generation.

Overall, after considering the authors' response and the contributions of the paper, I still find the methodological novelty insufficient. Therefore, I maintain my original score.

We thank the reviewer for the follow-up and for acknowledging both the practical importance of the problem and the effectiveness of the experimental results.

We understand the remaining concern regarding methodological novelty. We do not claim to introduce fundamentally new operators; wavelet decomposition, cross-attention, and gated injection are established techniques. Our contribution lies in the problem formulation and the specific design for garment consistency. Low-frequency structure establishes spatial correspondences that track the garment’s deformation, while high-frequency details are retrieved from a fixed reference bank and injected along those correspondences to enable consistent detail propagation under free motion generation.

We respect the reviewer’s assessment of novelty and leave the final judgment to the Area Chair. We hope the clarified problem setting, the controlled experiments, and the demonstrated gains in detail fidelity will be taken into account in the overall evaluation.

Reviewer_CFv4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses garment-detail drift in fashion videos generated from reference images and motion prompts. It proposes W-DiT, a framework based on Wan2.2 that uses separate reference pathways for structure and fine appearance details. As I understand it, the main idea is to separate where the model retrieves information from what information it retrieves. Haar wavelet decomposition produces low-frequency and high-frequency components from each reference image. The low-frequency component passes through the VAE to provide structural features, while a separate CNN encodes the high-frequency details. During denoising, the model matches low-pass-filtered DiT features against the reference structure. It then uses the attention weights to retrieve reference details and adds them to selected DiT blocks through gated residual updates. I view this as learned feature retrieval and spatial resampling. The reference detail bank stays fixed, while the retrieval weights change with the generated video. Both the compatibility of the two structural representations and the adaptation of details to garment deformation are learned through training. This provides a continuing source of reference appearance information, although it does not by itself guarantee precise geometric tracking of textures. The authors introduce FashionVideo, with approximately 10,000 synthesized videos and 30,000 associated images. Experiments include a matched-budget Wan2.2 comparison, ablations, evaluation on captured fashion videos, and human judgments. The reported results support better garment appearance and identity preservation with comparable motion quality. They provide clearer evidence for overall appearance consistency than for accurate local correspondence under complex deformation. The editing application uses target-outfit images prepared by an upstream image editor.

Critical strengths
  1. The method addresses a clear problem in fashion video generation: garment details tend to drift as the person moves. The authors propose to use structural features to decide where to retrieve information, and a separate high-frequency bank to supply the details, which is a reasonable design for this problem.
  2. I find the single-reference comparison with Wan2.2 useful, since the training budget and generation settings are matched. The DFS improvement from 72.1 to 82.7 is substantial. The RGB and standard-attention ablations also help explain the gains, while the results on captured fashion videos and the human evaluation provide further support for improved appearance consistency.
Critical weaknesses
  1. My main concern is whether the experiments support the paper’s interpretation of detail propagation. The reference structure comes from VAE-encoded low-pass images, while the target structure comes from low-pass-filtered DiT hidden features. Learned projections may make these representations compatible, but their geometric correspondence needs to be checked. Attention retrieval can redistribute features spatially, yet the paper does not directly measure whether individual details stay correctly attached to the garment through pose changes, deformation, and occlusion. DFS measures garment feature similarity, and the heatmaps mainly show that attention reaches the relevant region. These results are useful, but they do not establish accurate correspondence within a logo or texture pattern.
  2. Several details needed to reproduce the method are missing, including the training schedule, trainable backbone components, target low-pass operator, and the injection layers and locations within each block. The additional parameter, memory, and inference costs are also unclear. The ablations leave some questions about where the gains come from: the RGB comparison does not isolate the reference VAE bypass, and the standard-attention comparison changes both queries and keys. The editing experiment also lacks a measure of target-garment fidelity or edit correctness. These gaps make it difficult to reproduce the results and judge the contribution of each design choice.
Questions for authors
  1. How do you establish that the reference VAE features and the low-pass-filtered target hidden features provide compatible structural information? Could you evaluate the correspondence on garment or logo landmarks, beyond the region-level heatmaps?
  2. Can you separate the benefit of bypassing the reference VAE from the benefit of using high-frequency inputs? Likewise, can you isolate the effect of low-frequency reference keys, since the standard-attention ablation changes both queries and keys?
  3. Please specify the target low-pass operator, injection layer indices, locations within each block, parameter sharing across blocks, and which backbone components are trained. What training steps and learning rate are used, and how much parameter, memory, and inference overhead does the method add?
  4. Please clarify how DFS features are aggregated, how garment segmentation failures are handled, which frames are sampled, and how ColorAccuracy is calculated. Which baselines were fine-tuned, which versions were used, and which output settings could not be matched?
  5. Could you add a focused evaluation of the target garment in the editing and detail-preservation experiments? Possible measures include local perceptual comparisons after independent alignment, recognition accuracy or CER for readable text, and human judgments of logo shape and placement. A motion-compensated temporal evaluation would also help check whether the details remain stable during movement.
1 comment
AuthorsOpenReview ↗

Response to reviewer CFv4

[W1, Q1: Local correspondence] We are grateful for this insightful comment. To address whether the alignment supports fine-grained detail propagation, we add point-level attention visualizations and quantitative correspondence evaluation.

Qualitative evidence. Appendix B.7 (Point-Level Reference Attention) presents nine point-query attention maps across three garment patterns, showing that queries at different locations within the same pattern attend to their corresponding local reference regions under pose changes.

Quantitative evidence (Appendix B.6). For each of 100 videos, we manually pair five garment landmarks in the reference image and one clear generated frame. Adapting DiffTrack (NeurIPS 2025), we use CoTracker to track the video landmarks and flag visibility. At each sampled denoising step, we query the model’s attention at each visible landmark. Using the DFS protocol, Grounded-SAM automatically supplies the reference garment mask, whose tight bounding box has diagonal L = √(w²+h²). A match is correct when the Euclidean distance between the reference-attention peak and the annotated reference point is at most 0.10L, measured in the same reference-image coordinates. [email protected] reports the percentage of correct matches per video, averaged across videos for each stage.

MethodEarly: τ∈[0.7,1.0]Middle: τ∈(0.3,0.7)Late: τ∈[0.0,0.3]
HF + Standard Attention62.4%74.2%78.1%
W-DiT71.8%81.5%84.9%

The 9.4/7.3/6.8 percentage-point gains provide quantitative evidence of more accurate local correspondence across denoising stages, supporting structure-guided detail retrieval during garment deformation.

[W2, Q2: Controlled ablations] Thank you for suggesting these controlled ablations. Appendix B.4 separates detail-bank encoding from input frequency:

Encoding pathRGB DFS ↑DWT HF DFS ↑
Reference VAE77.878.5
Independent bypass encoder80.282.7

Appendix B.5 isolates the query and reference-key choices:

Query / reference keyDFS ↑[email protected] ↑
Unfiltered / HF78.472.6%
Low-pass / HF79.574.4%
Unfiltered / low-frequency structure80.676.5%
Low-pass / full-band RGB VAE81.477.8%
Low-pass / low-frequency structure82.779.7%

Changing only the query or key improves DFS by 1.1 or 2.2. Combining them yields +4.3 DFS and +7.1 PCK points. Against the full-band VAE key, the low-frequency key adds 1.3 DFS and 1.9 PCK points. Motion varies only from 94.2 to 94.6. Full appearance, motion, and identity results are included in Appendices B.4-B.5.

[W2, Q3: Implementation and overhead] Appendix B.8 provides the Gaussian low-pass configuration, injection blocks and placement, parameter sharing, trainable components, and training schedule. The backbone DiTs, VAE, and text encoder are frozen. Reference encoders are shared across blocks and experts. Alignment and injection modules are block- and expert-specific. Appendix B.3 reports 421M total / 221M active added parameters, memory, and 5.3-5.7% latency overhead.

[W2, Q4: Evaluation and baseline settings] DFS uses 16 uniformly sampled frames, mask-pooled DINOv2 garment features, best-reference matching, and video-level averaging. Appendix B.9 details feature extraction, segmentation fallbacks, and each baseline's checkpoint, additional inputs, and native output settings. Color Accuracy uses VBench's color metric. Appearance and Motion average their stated VBench components, and FaceID uses ArcFace. Our method and the Wan2.2 control are fine-tuned on FashionVideo. Other baselines use released weights/pipelines or APIs.

[W2, Q5: Local pattern fidelity] We add a focused human evaluation of 100 videos per method, directly assessing target-garment pattern shape, placement relative to the garment, and detail stability during movement (Appendix B.2):

MethodPattern shapePlacementStability during motionOverall pass
Wan2.238%52%34%29/100
W-DiT78%85%76%73/100

Motion-compensated quantitative evaluation (Appendix D.2). This is a helpful suggestion. Garment Ewarp measures RGB reconstruction error after external optical-flow alignment within valid, non-occluded garment regions. Trajectory-Aligned DFS measures DINOv2 cosine similarity between visible CoTracker trajectory points and their manually paired reference points:

MethodGarment Ewarp ↓ (×10⁻³)Trajectory-Aligned DFS ↑
Wan2.25.1870.3
W-DiT (1 Ref)3.4181.6
W-DiT (3 Refs)3.1884.2

With one reference, W-DiT reduces warping error by 34.2% and improves trajectory-aligned fidelity by 11.3 points. Three references further improve both measures. These results quantify motion-compensated temporal consistency and local reference fidelity, complementing the human judgments of pattern shape and placement. Appendix D.1 additionally reports segment-wise DFS.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their careful reading and constructive feedback. The comments have substantially improved the clarity and completeness of our work.

In response, we have made the following major revisions and additions:

  • Quantitative correspondence evaluation: We introduce a [email protected] protocol (adapting DiffTrack with CoTracker tracking and Grounded-SAM garment boxes) that measures whether structure-guided attention correctly retrieves local garment landmarks across denoising stages. Results show consistent gains of 6.8-9.4 percentage points over the HF + Standard Attention baseline (Appendix B.6). Point-level attention visualizations are also provided (Appendix B.7).

  • Controlled ablations: We separate the benefits of (i) bypassing the reference VAE versus using high-frequency inputs, and (ii) low-pass target queries versus low-frequency reference keys (Appendices B.4-B.5). These ablations confirm the complementary contributions of each design choice.

  • Local pattern fidelity and motion-compensated metrics: We add a focused human evaluation of pattern shape, placement, and stability during motion (Appendix B.2), together with Garment Ewarp and Trajectory-Aligned DFS (Appendix D.2).

  • Comparison with modern video models: Under a unified human usability protocol (garment detail + identity + reasonable motion), we compare W-DiT against Wan2.5, Veo 3, Omni, and Seedance 2.5 on 100 held-out reference-prompt pairs (Appendix B.1). W-DiT achieves substantially higher usability.

  • Implementation and evaluation details: Full specifications of the low-pass operator, injection blocks and placement, parameter sharing, training schedule, parameter/memory/latency overhead, DFS computation protocol (including segmentation fallback), and baseline settings are now documented (Appendices B.3, B.8, B.9).

We kindly ask the reviewers to examine the revised manuscript (especially the new Appendices B, C, and D) and the detailed point-by-point responses below. We believe these additions address the main concerns regarding correspondence accuracy, experimental rigor, novelty positioning, and the value of the proposed approach relative to modern general-purpose video models.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewers,

Thank you again for your valuable comments. As the discussion period is ending soon, we just wanted to check whether our clarifications and additional experiments have addressed your main concerns. We would be happy to provide any further clarification if needed.

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.