acceptodds
Under review as a conference paper at ICLR 2027

DualFashion: Dual-Sided Garment-Conditioned Multi-View Fashion Model Generation via Hierarchical Structured Inpainting

Abstract

Most virtual try-on methods edit a supplied full-body person image, while garment-driven virtual dressing typically generates a single image from a frontal garment reference. Neither setting directly supports creating a coherent multi-view fashion presentation from complementary garment evidence. We introduce DualFashion, a framework for dual-sided garment-conditioned multi-view virtual dressing. Given paired front and back garment photographs and a head reference, DualFashion directly generates an ordered set of natural-presentation, front, side, and back views, without requiring a target body image, pose, or scene. The task requires accurate routing of front- and back-side garment details, consistent garment and subject appearance across views, and sufficient spatial capacity for fine-grained details. We address these challenges through hierarchical structured inpainting. A shared-canvas joint stage first synthesizes a coordinated four-view set within a single denoising process. A set-conditioned view-wise stage then reconstructs each view at higher resolution using the complete intermediate set together with the original garment and head references, improving local details while preserving set-level coherence. To support this task, we introduce DVF-6K, a dataset of 6,011 product groups across 33 categories, pairing front and back retail garment photographs with natural-presentation, front, side, and back on-body views. Experiments demonstrate improved visual quality, garment fidelity, identity preservation, viewpoint correctness, and cross-view consistency. The generated views can also serve as sparse conditions for downstream fashion-video synthesis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-view Virtual Dressing, Dual-sided Garment Conditioning, Hierarchical Structured Inpainting, Shared-canvas Joint Generation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_qdYJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DualFashion, a new setting for dual-sided garment-conditioned multi-view virtual dressing. Given paired front/back garment photographs and a head reference, the system generates an ordered four-view fashion presentation consisting of natural-presentation, front, side, and back views, without requiring a target full-body image, pose, or scene.

Critical strengths

The strongest contribution is the formulation of dual-sided garment-conditioned multi-view virtual dressing together with a dataset whose supervision structure actually matches this task. DVF-6K explicitly links front/back garment evidence to ordered natural/front/side/back catalog views.

Critical weaknesses
  1. Magic Clothing, DreamFit, and IMAGDressing-v1 use their released checkpoints, one frontal garment image, and external FaceID interfaces, whereas DualFashion is specifically trained on DVF-6K using FLUX.1-Fill and receives both front and back garment photographs. Qwen receives the same two garment references, but is a generic image-editing model rather than a model trained for the proposed task. The paper therefore demonstrates that the complete task-specific DualFashion system outperforms these existing systems under the authors' protocol, but it does not cleanly establish how much of the gain comes from the hierarchical method itself versus task-specific training, a stronger/different backbone, paired garment evidence, or the new dataset.
  2. The Ablation Does Not Isolate the Key Mechanisms Claimed by the Paper. Table 4 compares only single-stage high-resolution joint generation against the complete DualFashion hierarchy. However, several central claims are more specific than what this ablation establishes. In particular, the manuscript argues that:
  • paired front/back evidence enables side-specific routing;
  • shared-canvas generation provides cross-view coordination;
  • role-specific embeddings improve semantic-view control;
  • auxiliary face/gradient/Laplacian objectives recover specific details.

The current experiments do not separately test these factors.

Questions for authors
  1. Can the authors ablate the central design factors separately—for example: front-only vs. front+back garment conditioning, Stage I only, Stage II conditioned only on the corresponding Stage-I view rather than the full set, and with/without the view router or auxiliary reconstruction losses?
  2. What is the end-to-end inference time and peak GPU memory for producing one four-view set on the stated single RTX PRO 6000 setup, and how does this compare with the single-stage high-resolution ablation?
  3. Can the authors provide a task-matched same-backbone baseline using FLUX.1-Fill and the same DVF-6K training data, but generating the four views independently or without the proposed hierarchy?
Reviewer_DbWyOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DualFashion, a framework for dual-sided garment-conditioned multi-view virtual dressing. Given paired front/back garment photographs and a head reference, the method directly generates an ordered set of natural-presentation, front, side, and back views without requiring a target full-body image, pose, or scene. Built on FLUX.1-Fill, the method adopts hierarchical structured inpainting: a shared-canvas joint stage coordinates the four views within a single denoising process, and a set-conditioned view-wise stage performs high-resolution reconstruction using the complete intermediate set and the original references. To support this task, the authors construct the DVF-6K dataset (6,011 product groups, 36,066 images, 33 categories), which uses product-level groups to associate paired retail garments with ordered semantic on-body views.

Critical strengths

The paper is relatively clear and well structured. The presentation is generally easy to follow, and the overall organization of the paper is reasonable.

The dataset construction is task-targeted. The construction of DVF-6K directly responds to the requirements of the proposed task.

Critical weaknesses

This paper lacks sufficient overall innovation, and its experimental validation of effectiveness and generalization is inadequate: Missing paired front/back multi-view try-on baselines. The comparison is unfair with respect to the multi-view capability that DualFashion claims. The first three baselines (Magic Clothing, DreamFit, and IMAGDressing-v1) receive only a frontal garment image, and the sole paired front/back baseline, Qwen-Image-Edit, is not a dedicated multi-view try-on method. Generalization validated on only one private dataset. All experiments use the authors' private DVF-6K dataset, and public datasets suitable for multi-view try-on (DeepFashion, MPV, and Eevee) are not used.

Questions for authors
  1. Insufficient comparative experiments: In the comparison, the first three baselines (Magic Clothing, DreamFit, IMAGDressing-v1) receive only a frontal garment image, while the only paired front/back multi-view try-on baseline is Qwen-Image-Edit, which is not a dedicated multi-view try-on method. To my knowledge, several recent works take paired front/back garments as input for multi-view try-on, such as Robust-MVTON (CVPR 2025), MV-VTON (AAAI 2025), and VC-VTON (TIP 2026). The authors should add such paired front/back multi-view try-on baselines to verify DualFashion's generation capability in multi-view try-on.

  2. Insufficient generalization validation: All experiments are conducted solely on the authors' private DVF-6K dataset. To my knowledge, other datasets applicable to multi-view try-on include DeepFashion, MPV, and Eevee. Although these datasets do not simultaneously provide flattened front and back garment views, they all contain images of the same model wearing the same garment from more than four viewpoints—especially Eevee. The authors could extract garment front and garment back from these data to validate try-on results from other viewpoints.

  3. Missing quantitative evaluation for downstream video synthesis: Section 5.4 presents only qualitative results for sparse-anchor video synthesis (Figure 5), without any quantitative metrics (e.g., inter-frame consistency, fidelity at anchor points, temporal smoothness). Since the paper lists "generated views can serve as downstream conditions" as one of its contributions, a few qualitative frames are insufficient to support this claim. Moreover, the generated video frames in Figure 5 appear to be of poor quality.

  4. Head-conditioning not visually verifiable in multi-view outputs: This task explicitly conditions on a head reference; however, in Figures 3, 4, 6, and 7 (as well as Figure 7 in the supplementary material), the generated front, side, and back views are all cropped above the neck. Since the head reference is a core input and the paper claims improved identity preservation and cross-view consistency, readers cannot visually verify whether head/identity information is actually routed to these views. Please clarify whether this cropping is an intentional dataset convention (i.e., garment-focused framing in DVF-6K) or a limitation of the model. If it is intentional, please provide visualizations showing full-body outputs with consistent heads across all four views to substantiate the identity-preservation claim. If the side/back ground-truth images in DVF-6K are inherently headless due to cropping, we suggest validating DualFashion on the DeepFashion dataset, which provides complete displays of clothing and model heads across multiple views, including front, back, and side. We therefore suggest validating DualFashion on DeepFashion to determine whether it can preserve identity in multi-view generation for multi-view virtual try-on.

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

This paper proposes DualFashion, a two-stage framework for generating consistent multi-view fashion images from front and back garment photographs and a head reference. It first jointly generates four views using shared-canvas inpainting, then refines each view at higher resolution with set-conditioned reconstruction. The authors also introduce DVF-6K, a dataset of 6,011 product groups across 33 categories. Experiments demonstrate improvements in garment fidelity, identity preservation, visual quality, and viewpoint correctness, supported by ablation studies and a user study.

Critical strengths

None

Critical weaknesses

1.Limited visual realism and multi-view consistency. The generated images in Figure 3 still exhibit noticeable AI-generated artifacts and unnatural appearances, limiting their practical value. Moreover, the evaluation lacks direct measurements of cross-view structural consistency.

2.Limited novelty and insufficient ablations. The method mainly combines existing inpainting and refinement techniques. Key ablations are missing, making it difficult to identify the actual contributions of each component to the reported improvements.

Questions for authors

1.Visual realism: How do the authors explain the noticeable synthetic appearances and unnatural human poses in Figure 3, despite the reported improvements in image quality metrics?

2.Ablation studies: Can the authors provide controlled experiments to separately evaluate the contributions of dual-sided garment conditioning, joint generation, and set-conditioned refinement?

3.Multi-view consistency: Can the authors provide quantitative evidence of cross-view structural consistency beyond CLIP and DINO similarity metrics?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.