- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
LayeredFashion decomposes a clothed-person image and outfit description into editable garment RGBA layers and a person underlay. It reverses the specified dressing order through alternating garment extraction and image restoration. An RGB-Semantic Alpha Refiner (RSAR) corrects garment opacity, while a separate cumulative repair region controls recurrent image updates. The authors construct 22,431 outfits with 97,775 garment layers and evaluate decomposition, text-to-PSD composition, and downstream 3D reuse. The method achieves the best reported means across twelve decomposition metrics.
Critical strengths
- Independently editable garment layers support recomposition and asset reuse beyond flat virtual try-on or visible-region human parsing.
- Separating saved garment opacity from the cumulative restoration region allows earlier residuals to be repaired without enlarging the current garment layer.
- Seven ablations examine alpha decoding, RGB and semantic evidence, full-image correction, and recurrent updates; removing RSAR substantially increases alpha error and trajectory drift.
Critical weaknesses
-
Performance on real clothing photographs is not sufficiently demonstrated. Quantitative evaluation mainly uses independently generated garments that are segmented and composited with source-over blending. This construction does not reproduce cross-layer shadows, physical fabric interactions, or complex real-world occlusion. The reference person underlay may itself retain clothing. Strong synthetic results therefore do not establish complete garment recovery from real photographs.
-
Generalization to complex clothing combinations remains uncertain. The dataset uses a closed prompt pool, explicitly specified categories and order, category-based coordinate priors, and specific text-processing rules. Although the representation distinguishes inner tops from outerwear, it does not demonstrate reliable separation of multiple overlapping garments within the same category, which may instead be segmented as a single layer. This limitation may reflect both restricted training coverage and the lack of explicit completion of individual extracted garments. Tests with more complex upper-body layering, free-form descriptions, and unfamiliar garment combinations are needed.
-
The claim of complete garment recovery is insufficiently supported. Each extracted garment combines the current image’s RGB with a refined alpha matte, without an explicit appearance-completion step for that garment. The upper-body garment layers illustrated in Figures 2 and 4 appear incomplete, particularly around occluded or rear portions.
-
The dataset could be a valuable resource for the research community, but the paper does not explicitly commit to releasing it. Clarifying the release plan would strengthen the dataset contribution and support reproducibility.
Questions for authors
Please refer to Weakness 2 3 and 4. If the dataset will be open-source upon the acceptance of this paper. I'd like to considering raising my socre.
5 comments
Response to Reviewer JiTb (Part 2 / 2)
W3. The claim of complete garment recovery is insufficiently supported. Why do Figures 2 and 4 (now Figures 3 and 5) look incomplete?
We thank the reviewer for highlighting this ambiguity in our original wording. By “complete garment recovery,” we mean recovering the garment regions that should be visible from the input viewpoint after overlying garments are removed, while preserving the occlusion relationships required for PSD-style recomposition. This does not imply recovery of the garment’s full 360° appearance or physical geometry. We have revised the manuscript to define “complete” precisely.
(1) What is completed, and how. Our completion target is the garment appearance that should be visible from the input viewpoint after overlying garments have been removed. This includes regions occluded by outer garments in the original image, but excludes regions that remain hidden behind the body, such as rear panels behind the torso or collar surfaces hidden by the neck.
Restoration explicitly completes the garment regions exposed by removing overlying garments, before those garments are extracted. In Figure 1, for example, it reconstructs the white shirt’s previously hidden front and sleeves, as well as the upper portions of the trousers covered by the jacket. Subsequent Extraction uses this restored RGB together with the refined alpha matte. The recovered layers therefore include appearance that was occluded in the original input.
(2) Why some regions remain transparent? This is an intentional design choice consistent with our paradigm. We follow PSD-style compositing and the physical occlusion order in the input view. For the open jackets in Figures 3 and 5, rear panels behind the torso are excluded from the RGBA layer, as are collar surfaces hidden by the neck. Filling these areas would cause the garment layer to obscure the body or inner clothing that should remain visible when recomposed. To further clarify the distinction between PSD-style layer extraction and full-garment extraction, we have added Figure 2 (Section 1, p. 2), which presents each isolated coat layer alongside its recomposition over the same underlying layers. In the left pair, the PSD-style layer leaves the coat opening transparent, so the dark inner garment remains visible. In the right pair, directly overlaying the full-garment layer covers the inner garment with the completed rear panel. Our core target is therefore a recomposable, view-aligned clothing layer, rather than a reconstruction of the garment’s full physical geometry.
The recovered layers can support further completion toward physically complete garment representations, as illustrated by our downstream 3D reconstruction results in Figure 1 and Figure 9. However, this serves a different objective from PSD-style decomposition, which requires preserving the visibility and occlusion relationships of the input view.
In response to your concern, we have clarified the scope of “complete” and the intended PSD representation in Sections 1 and 3.1 and Appendix A.1, with corresponding revisions to the text and supporting figures.
W4. Dataset release: will the dataset be made available upon acceptance?
We commit to publicly releasing the complete LayeredFashion dataset, including all layer annotations and text descriptions, upon acceptance to support reproducibility and further research. This commitment is now explicitly stated in the Reproducibility statement (p. 11).
We sincerely hope that our response has addressed your concerns. Your thoughtful and perceptive feedback is sincerely appreciated and has genuinely helped us further strengthen the manuscript. Should you have any further questions, please let us know. We would be more than happy to address them and make further improvements.
Thank you for the revisions. The paper has improved since the initial submission, and the commitment to publicly release the dataset is a significant contribution. After reading the new material, I still have several concerns.
First, missing regions of outer garments and fabric occluded behind the body remain a substantial issue, as illustrated by the scarf in Figure 1, the jackets in Figures 4 and 7, and the hat in Figure 7. The claimed “complete garment recovery” appears limited to partial recovery of inner garments.
Second, all examples shown are strictly front-facing and feature no garment overlap or occlusion. Given how the method works, occlusion by the arms, such as when they are crossed in front of the body, or a side-facing pose could substantially affect decomposition quality.
Third, the real-world evaluation consists of only a few qualitative examples, which are insufficient to establish generalization in practice. Training and testing on a larger set of real photographs would strengthen the work.
For these reasons, I maintain my original Weak Reject rating.
Response to Reviewer JiTb (Part 1 / 2)
Thank you very much for your thoughtful and perceptive feedback, which has genuinely helped us further strengthen the manuscript.
W1. Real photographs: do synthetic composites capture shadows, fabric interactions, and complex occlusion; why does the underlay retain clothing?
Thank you for raising this important question. Your insights have deepened our understanding of our own work and have been particularly valuable in strengthening this work. While conducting the additional experiments you suggested, we were also pleased to discover some capabilities and generalization of our method that we had not anticipated.
(1) Real-world evaluation. Following your suggestion, we evaluated our method on real-world photographs and challenging examples involving fabric interactions (e.g., complex wrinkling in layered garments) and extreme inter-layer occlusion in Section 6.3, Figures 6 and 7. The results show that our method can not only accurately recover the layers for real-world images, but also reconstructs physically plausible wrinkles in occluded regions, with patterns consistent with pose-induced stretching. These results provide evidence beyond source-over synthetic composites.
(2) Underlay. The underlay is intentionally retained on the base body to protect bodily privacy and avoid exposing intimate regions. By default, it is excluded from target garments and not extracted as a separate layer. Its presence reflects the intended decomposition scope, rather than incomplete segmentation. We clarified this in Section 3.1 and Appendix A.1.
(3) Additional poses. We further tested a series of real captured images, more varied poses, multiple layers of similar clothing, and their combinations. These experimental results have been updated in Section 6.3, Figures 6 and 7. They demonstrate the robustness and generalization of our method to real and complex inputs.
W2. Generalization: Generalization to complex clothing combinations remains uncertain; tests with multiple overlapping garments within the same category, and free-form descriptions are needed.
Thank you for raising this concern. We have clarified the role of these design choices and added real-world examples with more complex layering and varied descriptions in Figures 6 and 7 in Section 6.3.
(1) Training structure and inference flexibility. The closed prompt pool and explicit category/order annotations are used to construct synthetic training data and provide consistent supervision. At inference, the parser derives garment categories and layer order from the supplied description, and the text-processing rules standardize this information for conditioning. Users can therefore vary the wording and the order in which garments are mentioned, as illustrated by the examples below.
The category-based coordinate priors provide coarse spatial guidance for garment localization. Localization and alpha prediction also depend on the input image, so these priors guide extraction without prescribing fixed garment masks. Inference thus retains structured conditioning while allowing descriptions beyond the training templates and using image evidence to determine the extracted layers. We have clarified this distinction in Section 4.1. Appendix E describes the parser, while Section 4.2 and Appendix C.1 detail inference conditioning.
(2) Same-category overlap. We added real-world examples in Section 6.3 with multiple overlapping T-shirts and multiple overlapping suit jackets. In these examples, the garments are successfully recovered as separate editable layers, demonstrating same-category separation in the tested cases. The scope of garment completion is discussed in our response to W3.
(3) Free-form descriptions. We invite the reviewer to examine the text prompts in the following table for results in Figures 6 and 7. These prompts vary in wording, garment-mention order, and natural sentence structure. Specifically, we can introduce clothing in different ways and can mention lower-body garments or accessories before upper-body layers. The corresponding outputs retain separate editable layers, providing qualitative evidence that the method can process these forms without requiring a fixed sentence template or garment-mention order.