CycleVTON: Improving Diffusion-Based Virtual Try-On with Cycle-Consistent Training

1. The explanation of the methodology could benefit from clearer clarification, as it is difficult to understand which phrase refers to which "image" in the pipeline. For instance, in lines 261-267, adding notations and labeling components in Figure 2 would help avoid ambiguity. Readers may struggle to differentiate between the frontal-view image and the distorted image and add labels (e.g., A, B, C) to each image in Figure 2 and use these labels consistently in the text description will be helpful. Additionally, in Figure 2, using different clothing items for the try-on and take-off procedures would help reduce confusion, as presenting the same blue t-shirt makes the process harder to follow.

2. The experiments were conducted on a dataset with professional fashion images, but in real-world virtual try-on applications, "in the wild" images are typically used. However, there is no experiment demonstrating how the proposed method performs on in-the-wild images. Could this suggest that the model may be overfitting to the fashion dataset and might struggle to adapt to more diverse, real-world images? the authors could conduct additional experiments on a dataset of non-professional, real-world images to demonstrate the model's generalization capabilities.

3. The poses in the presented dataset appear to involve relatively limited movement. It would be beneficial to include examples or conduct additional experiments with more extreme or dynamic poses to demonstrate the model's robustness.

4. L112 typo: ‘arping’ →’Warping’

1. Significant clothing deformations or angle changes lead to artifacts such as wrinkling or stretching, which limit the model's ability to accurately reproduce the original textures and patterns.

2. It depends on the standardized front views for optimal performance, which may limit its effectiveness with certain types of clothing.

3. The dual-network cycle-consistent approach may raise computational costs and complexity, potentially affecting its use in real-time or resource-limited environments.

4. The cycle idea is very close to this paper, so the novelty needs to be clarified. https://openaccess.thecvf.com/content/CVPR2021/papers/Ge_Disentangled_Cycle_Consistency_for_Highly-Realistic_Virtual_Try-On_CVPR_2021_paper.pdf

1. In the proposed architecture, the CLOTHING EXTRACTION NETWORK differs only in its function, yet it does not fundamentally diverge from the ReferenceNet used in previous methods, such as IDM-VTON.

2. The proposed Cycle-Consistency Optimization Strategy is successful. However, there is a lack of validation for the effectiveness of its subprocesses and sub-losses, such as L^_h and L^_c.

3. Some of the data in Table 1 seem to differ from those in the original paper, why?

4. Please provide the code to demonstrate reproducibility.

5. The entire process is overly engineered, lacking rigorous theoretical validation in this type of conference.

The idea of extracting human feature and clothing feature in a cycle framework is not novel, i.e., a similar work CycleVTON: A Cycle Mapping Framework for Parser-Free Virtual Try-On [Du et al., AAAI 2024]. The paper did not cite or discuss the CycleVTON paper [Du et al.]. A comparison against CycleVTON [Du et al.] should be provided since the ideas of the two papers are similar.

The technical contribution is limited. The virtual try-on part (ReferenceNet and denoising UNet) is borrowed from Animate Anyone [Hu et al.], and the main contribution is the idea of a clothing extraction network, which, however, has appeared in a previously published paper CycleVTON [Du et al.]. The technical contributions should highlight the differences against existing works.

No discussion on the failure cases (such as particular clothing types and poses). What are the limitations of the method?