acceptodds
Under review as a conference paper at ICLR 2027

LiWi: Layering in the Wild

13 views

Abstract

Recent advances in generative models have empowered impressive layered image generation, yet their success is largely confined to graphic design domains. The layering of in-the-wild images remains an underexplored problem, limiting fine-grained editing and applications of images in real-world scenarios. Specifically, challenges remain in scalable layered data and the modeling of object interaction in natural images, such as illumination effects and structural boundary. To address these bottlenecks, we propose a novel framework for high-fidelity natural image decomposition. First, we introduce an *Agent-driven Data Decomposition* (ADD) pipeline that orchestrates agents and tools to synthesize layered data without manual intervention. Utilizing this pipeline, we construct a large-scale dataset, named **LiWi-100k**, with over 100,000 high-quality layered in-the-wild images. Second, we present a novel framework that jointly improves photometric fidelity and alpha boundary accuracy. Specifically, photometric residual prediction models differences unexplained by RGBA compositing, and degradation-restoration objective provides boundary-correction supervision by recovering clean foreground image from degraded one. Extensive experiments demonstrate that our framework achieves state-of-the-art (SoTA) performance in natural image decomposition, outperforming existing models in RGB L1 and Alpha IoU metrics. We will soon release our code and dataset.

Primary area:
generative models
Keywords:
Image Decomposition, Image Layering, Diffusion Models

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_DvqjOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents a novel approach to decomposing an image into a background layer and foreground layers. A main bottleneck in natural image decomposition is the lack of training data. To overcome this, the paper proposes a dataset synthesis pipeline that avoids manual labeling, a dataset generated from in-the-wild images using the proposed dataset synthesis pipeline, and a layer decomposition framework with photometric residual prediction and auxiliary layer refinement. The proposed pipeline consists of multiple steps, which perform different subtasks, e.g., foreground detection and removal for background layer extraction, foreground extraction, layer composition and proposal selection, and verification. For performing subtasks, the proposed pipeline adopts agents and specialized tools. The resulting dataset includes a source image, a background layer, and a set of foreground layers and their alpha masks. The layer decomposition framework is built upon a pretrained diffusion-based layer decomposition model, Qwen-Image-Layered. To better model complex photometric effects in real-world photographs, such as cast shadows, illumination variations, and contact darkening, the paper also proposes a photometric residual layer that captures disparities between the source image, and the reconstruction result where all synthesized layers are combined. The paper also proposes degraded boundary refinement, a training scheme that augments flow-matching training paths to explicitly include potential boundary degradations. Experimental results show that the proposed method achieves high-quality layer decomposition results.

Critical strengths

The LiWi-100k dataset can be valuable to the community. It offers a large collection of layered images of diverse scenes with multiple foreground layers.

The dataset synthesis pipeline seems to be carefully designed, as it achieves 94.9% human acceptance rate.

Modeling photometric effects using a separate layer is also an interesting idea. Previous methods often include them in foreground layers, but such photometric effects are often results of interactions between foreground and background objects.

The proposed degraded boundary refinement scheme is also reasonably designed.

Critical weaknesses

Unfortunately, the paper has several weaknesses regarding its technical contributions, validation, and presentation.

  1. Unclear technical contribution of the ADD pipeline

The technical contribution of the proposed ADD is rather unclear. While the paper presents as an agentic framework, its overall workflow appears to be largely predefined, and agents are employed to perform specific subtasks such as foreground identification, segmentation, and result verification. Thus the proposed pipeline does not seem to be fundamentally differnet from conventional dataset synthesis pipelines with specialized tools. It is unclear what additional capabilities the proposed agentic design provides.

  1. Insufficient valdiation of ADD

The ADD pipeline consists of multiple components, each of which may introduce errors. However, the paper provides little analysis of the performance and reliability of these individual components. For example, it is unclear how the pipeline handles challenging cases such as semi-transparent objects and complex foreground boundaries present in natural images. Although the authors report a high human acceptance rate for the resulting dataset, this only measures the quality of the retained samples and does not reveal the failure rates or converage of the overall pipeline. Together with the unclear technical novelty discussed above, the lack of thorough validation makes it difficult to regard ADD as a technically solid contribution.

  1. Insufficient evaluation of the proposed dataset

Although LiWi-100k is presented as a major contribution, the paper lacks a systematic comparison with existing layered image datasets. In particular, comparisons of dataset scale, diversity, and layer complexity and quality would help understand the significance of the proposed dataset. Furthermore, it is not analyzed the coverge of the challenging cases such as semi-transparent objects, complex occlusions, and soft boundaries.

  1. Unclear motivation and interpretation of the photometric residual layer

The proposed photometric residual layer is not convincing. Fig. 4 shows that the residual layers have substantial high-frequency details and structural information that are difficult to interpret as photometric effects caused by interactions between foreground and background layers. Rather, these residuals appear to capture errors or missing details introduced by the decomposition process. Since the residual is defined simply as the difference between the source image and the recomposed image, it seems to include all kinds of decomposition errors, and the claim that the residual layer explicitly models photometric interactions does not look adequate.

The discussion in Section 4.1 does not sound convincing either. The authors suggest that high attention weights toward the clean source image may indicate information leakage, but this interpretation is speculative. Since the source image is explicitly provided as a condition, attending to it is not inherently problematic. Furthermore, it is unclear why a more balanced attention distribution should imply better decompostion or reduced information leakage.

  1. Insufficient description of the LiWi framework

Section 4 introduces the photometric residual learning and degraded boundary refinement, but does not clearly describe the overall LiWi architecture and its input and output. Although the paper mentions that the proposed framework is built using Qwen-Image-Layered, it is unclear how the proposed components are integrated into the original architecture.

  1. Insufficient clarity and qualitative analysis of the experimental results The experimental section is also difficult to follow. For example, the paper describes RGB L1 as measuring RGB reconstruction accuracy, but the paper does not clearly explain whether this metric evaluates individual foreground layers, background layers, or their combinations.

Furthermore the paper lacks qualitative abalation studies of the proposed photometric residual prediction and degradation-restoration objective. Such comparisons would be particularly important as there can be artifacts and characterstics that quantitative results cannot capture.

Questions for authors

Please address the weaknesses mentioned above.

3 comments
AuthorsOpenReview ↗

Part 1: W1-W2

Dear Reviewer,

Thank you for your careful review. We added stage-wise ADD statistics and dataset comparisons, and clarified the agent–verifier feedback loop and the motivation for residual prediction.

1. Unclear technical contribution of the ADD pipeline

Conventional pipelines fix objects/instructions and use manual review. ADD agents autonomously identify scene FG entities or specific BG content, generate adaptive instructions, and decide when to stop. The verifier replaces manual review, checking FG completeness, BG cleanliness, artifacts, and compositional plausibility. It flags specific issues for agent reprocessing: a processing–verification–correction feedback loop. Agents choose actions from scenes and feedback, beyond tool calls. We will detail this in the revision.

MuLAn-style [1] mask-based segmentation/inpainting can miss shadows/reflections outside masks eg˙,˙aremovedcup’stabletopshadowe\.g\., a removed cup’s tabletop shadow. We also observed mask shapes cause object regeneration and BG restoration failure: information leakage through the mask shape. ADD’s scene-specific agent instructions and verifier checks enable a closed-loop pipeline for constructing layered data from real-world images.

2. Insufficient validation of ADD

2.1 Component reliability and the verifier’s role

Identification, completion, FG extraction, and alpha estimation can fail, especially at semi-transparent regions/complex boundaries. The verifier rejects these errors (Question 1): Appendix A’s three independent annotators found verification raised pass rates from 52.7% to 94.9%, reducing errors in retained data.

Experiments in the appendix have already validated the reliability of some modules. Appendix H: DINOv2 [2] improves Proposal Selector human-audited pass rates from 43% pixel−wisedistancepixel-wise distance to 51%. Appendix G: VLM backbones without verification yield 51%, 47%, and 40%.

2.2 All pipeline: failure rate and source-image coverage

ADD stage-wise statistics for 5,000 source images are below.

StageInput count and unitOutput count and unitStage pass rateMain failure reasons
BIC: foreground identification and background generation5,000 source images8,742 BG candidates from 4,590 source images91.80% 4,590/5,000;persourceimage4,590 / 5,000; per source imageIncomplete entity identification; removal failure; abnormal background generation
FIC: foreground extraction and alpha fusion13,680 entity extraction tasks11,696 FG candidates85.50% 11,696/13,680;pertask11,696 / 13,680; per taskUnstable alpha in semi-transparent regions; missing thin structures; incomplete extraction of foregrounds or attached parts
Deduplication and Proposal Selector36,420 enumerated candidate combinations16,171 retained candidate combinations44.40% 16,171/36,420;percombination16,171 / 36,420; per combinationDuplicate combinations; redundant content across foregrounds; foreground remnants in the background; insufficient global consistency
VLM Verifier16,171 candidate combinations6,630 layered samples41.00% 6,630/16,171;percombination6,630 / 16,171; per combinationLocal artifacts; incomplete foregrounds; remaining objects or shadows in the background; implausible occlusion completion
All pipeline5,000 source images2,143 source images with at least one qualified output42.86% 2,143/5,000;persourceimage2,143 / 5,000; per source image2,857 source images produced no qualified output, including failures at individual stages and cases where all candidates were rejected
AuthorsOpenReview ↗

Part2: W3-W4

3. Insufficient evaluation of the proposed dataset

3.1 Comparison with existing datasets in scale, diversity, and layer complexity

The table below compares dataset scale, layers, image types, and shadows/natural interactions.

DatasetSource and scaleLayer structureImage typeShadows / natural interactions
Crello [3, 4]Design templates; 23,301 templates (LayerD split)Multiple design-element layers: text, vector graphics, and image assets; variable layer counts; element images, stacking order, and opacity are provided.Graphic designsComposited design layers; no scene-illumination supervision after object removal.
LAION-L²I [5]LAION-Aesthetics; 57.02M samples; 512×512Two fixed layers: 1 BG + 1 FG + mask; multiple salient objects share one FG and are not separated by instance.<br>Mixed web imagesCompletes the background occluded by the foreground; does not separately model object-related shadows/reflections.
MuLAn [1]COCO + LAION; 44,860 images; 101,269 instances2–6 layers: 1 BG + 1–5 instance FGs; depth/occlusion ordering; continuous alpha from VitMatte.Photographs and web imagesCompletes backgrounds and occluded instances; does not explicitly handle object-related shadows; additionally trains failure filters using human annotations.
LiWi-100k (ours)Constructed from natural-scene source images using ADD; 100,371 samples; 6 broad categories / 16 fine categories2–5 layers: 1 BG + 1–4 FGs; separate RGB and fused alpha; 55% have 2 layers and 45% have 3–5 layers.Natural scenesOcclusion completion and checks for removal of objects and associated shadows; VLM quality verification without separately training a filter; R represents the remaining difference between the source image and the RGBA composition.

BG = background; FG = foreground. Layer counts include BG but exclude R.

3.2 Coverage of challenging cases

Challenging cases include:

11 Lighting/shadows: Figure 11’s third example row from the top groups the person by the window, chair, and round table in Layer 1 and indoor BG in Layer 4, preserving natural lighting and handling object–floor shadows.

22 Occlusion: Figure 10’s third example row on the right separates person, wooden railing, and lake BG into three layers: complete railing behind person, then lake/shore behind railing.

33 Overlap/order: Figure 11’s bottom example has five layers—cake, wrapping paper, plate, tray, and BG—with occlusion completion and preserved stacking order. Its first-row dessert also has overlapping objects at different depths.

4. Unclear motivation and interpretation of the photometric residual layer

4.1 What does the residual capture, and is it limited to physical lighting interactions?

R = I_src − CompositeB,{Fk,αk}B, \{F_k, α_k\} comprises unexplained RGBA differences: shadows, illumination, and possible editing/alignment errors. We will clarify its definition and Figure 4: R is not a uniquely separated, purely physical illumination component.

R captures object-induced traces missing from compositing objects with cleaned BGs. Replacing source reconstruction with R also breaks the identity between target and clean input, so copying I_src cannot solve the auxiliary task. Table 4’s pixel-space residual ablation supports both motivations resultsin42˙results in 4\.2.

4.2 Why can source-image reconstruction create a shortcut, and is uniform attention required?

FG/BG attention to clean I_src preserves content. Reconstructing I_src instead permits copying the identical clean input, solving reconstruction without improving decomposition.

QIL [6] adopts ART’s [7] reconstruction target. ART has no I_src input in text-to-layer generation; predicting it anchors layer consistency. Text-image-to-layer generation already receives clean I_src, and Figure 5 toptop shows reconstructed I_src strongly attending to it, enabling copying without improving consistency.

Predicting R uses more layer information, consistent with I_src = CompositeB,{Fk,αk}B, \{F_k, α_k\} + R. We do not require attention to be uniformly distributed; instead, we want the model to use decomposition-related information. The ablation in Table 4 further demonstrates this point: removing reconstruction improves RGB L1/alpha soft IoU from 0.0996/0.8745 to 0.0950/0.9177; pixel-space R further gives 0.0910/0.9243.

AuthorsOpenReview ↗

Part3: W5-W6

5. Insufficient description of the LiWi framework

LiWi uses Qwen-Image-Layered [6]. Its four training inputs are clean I_src, noisy R ε→Rε → R, noisy BG/FG layers ε→x(layerε → x(layer), and an auxiliary degraded-FG path xd(layerx_d(layer + ε → x(layer)).

Auxiliary and regular paths share weights. Each degraded FG shares its target FG’s positional encoding for spatial alignment and attends only to itself and I_src, retaining source context while restoring eroded/dilated/blurred boundaries.

Inference removes degraded-FG inputs/path, retaining clean I_src and noise-to-residual/layer generation. Shared weights transfer learned boundary restoration without extra inputs or a restoration step. The revised diagram will show the four training inputs and simplified inference.

6. Insufficient clarity and qualitative analysis of the experimental results

6.1 What does RGB L1 evaluate, how does Max Edits work, and how are scores aggregated?

We follow LayerD [4] (Appendix B):

Max Edits = 0: group FG/BG layers by visibility, match predictions to references, and compute RGB L1 per pair without extra adjacent merges.

Max Edits > 0: after grouping/matching, allow up to that many additional adjacent-layer merges atmostoneforMaxEdits=1at most one for Max Edits = 1. Evaluate merged and unmerged layers.

Evaluation covers FG, BG, and partial compositions. Average scores over matched pairs per image, then over test images.

6.2 Qualitative ablations of residual prediction and degradation-restoration

We will add main-paper/appendix qualitative layer-output comparisons before/after removing source reconstruction, adding pixel-space R, and adding degradation-restoration, plus failure cases showing remaining limitations.

References

[1] Petru-Daniel Tudosiu et al.. MULAN: A Multi Layer Annotated Dataset for Controllable Text-to-Image Generation. CVPR, pp. 22413–22422, 2024.

[2] Maxime Oquab et al.. DINOv2: Learning Robust Visual Features without Supervision. Transactions on Machine Learning Research, 2024.

[3] Kota Yamaguchi (CyberAgent). Crello dataset. Hugging Face dataset card. Accessed November 10, 2026.

[4] Tomoyuki Suzuki, Kang-Jun Liu, Naoto Inoue, and Kota Yamaguchi. LayerD: Decomposing Raster Graphic Designs into Layers. ICCV, pp. 17783–17792, 2025.

[5] Xinyang Zhang, Wentian Zhao, Xin Lu, and Jeff Chien. Text2Layer: Layered Image Generation using Latent Diffusion Model. arXiv:2307.09781, 2023.

[6] Shengming Yin et al.. Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition. arXiv:2512.15603, 2025. Official model weights.

[7] Yifan Pu et al.. ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation. CVPR, pp. 7952–7962, 2025.

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

This work focuses on layered image decomposition for in-the-wild images and makes two main contributions: 1. It constructs a data pipeline for image layering, consisting of background, foreground, and layered image processing modules that incorporate several tools and models for different purposes. 2. It introduces two key methods to improve decomposition performance: an additional loss to capture high-level information, such as lighting effects, and an adapted rectified flow matching loss. Finally, this work achieves a new state-of-the-art result on a public benchmark.

Critical strengths

This work focuses on layered image decomposition for in-the-wild images and makes two main contributions: 1. It constructs a data pipeline for image layering, consisting of background, foreground, and layered image processing modules that incorporate several tools and models for different purposes. 2. It introduces two key methods to improve decomposition performance: an additional loss to capture high-level information, such as lighting effects, and an adapted rectified flow matching loss. Finally, this work achieves a new state-of-the-art result on a public benchmark.

Critical weaknesses

The writing needs improvement. For example, the baseline is not clearly identified until the experimental section, and there is no method diagram illustrating the residual design, making it difficult to understand what this work is based on. Some figures are also unclear; for example, the meanings of Layer 1 and Layer 2 in Figure 5 are not well explained.

Questions for authors
  1. How is the quality of the generated data evaluated? How does the verifier work?
  2. Where do the in-the-wild images come from?
  3. How do you ensure that the FG and BG are successfully separated? Currently, the verifier does not seem to check whether the FG is still contained in the BG.
  4. What is the verifier? It seems to perform many tasks, as described in L229–L234, but its details are not explained.
  5. Is I_c generated from the predicted B and F? Is I_src the ground truth? A diagram illustrating the method’s framework is missing.
  6. What do Layers 1 and 2 mean in Figure 5? How are these attention weights obtained? Are they averaged?
1 comment
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

Thank you for your helpful comments. We clarify the data-quality checks, verifier implementation, and LiWi’s training and inference below, and will improve the framework diagram and figure descriptions.

1. How is the quality of the generated data evaluated? How does the verifier work?

1.1 How is data quality evaluated?

Quality assessment combines automated checks and independent human audits. The Proposal Selector checks FG redundancy, FG remnants in BG, and global consistency; Qwen3-VL-32B [1] then verifies visual quality. Appendix A’s three independent annotators found pass rates of 52.7% without verification and 94.9% with it.

1.2 Verifier inputs, outputs, and feedback

Verifier inputs are I_src, all candidate layers, and a verifier skill. Outputs are is_valid true/falsetrue/false and explanation findings/reasonsfromBG,FG,andrecompositionchecksfindings/reasons from BG, FG, and recomposition checks. Failed samples are withheld; identified issues return to the agent for reprocessing.

2. Where do the in-the-wild images come from?

Inputs are unstructured natural-scene RGB images from the internal platform and can be released as open-source data.

3. How do you ensure that FG and BG are successfully separated? Does the verifier check FG remnants in BG?

We generate FG/BG candidates via object removal, BG completion, FG extraction, and multi-model alpha fusion, then select compatible combinations. Removal includes object-related traces such as shadows.

We explicitly check whether the background still contains the foreground. The FG–BG check compares DINOv2 [2] features within the FG-mask region; high FG/BG similarity indicates possible FG remnants and triggers rejection. The VLM verifier checks BG cleanliness and FG completeness, including remaining objects/shadows. Inter-FG checks remove duplicate content; Global Consistency checks agreement with the source scene.

4. What is the verifier, and how does it perform the checks in L229–L234?

Qwen3-VL-32B [1] checks BG cleanliness, FG completeness, and recomposition plausibility, complementing the Proposal Selector’s conflict/duplication checks inputs/outputs:Question1inputs/outputs: Question 1. Appendix A’s independent audit reports 52.7% pass rates before filtering versus 94.9% for retained samples. We will document the skill, visual-input organization, output fields, and use of decisions/explanations for filtering and feedback.

5. Is I_c generated from predicted B and F? Is I_src the ground truth? A framework diagram is missing.

5.1 Definitions of I_c, I_src, and the residual target

The dataset provides reference I_src and ADD-constructed B and {F_k, α_k}. I_c = CompositeB,{Fk,αk}B, \{F_k, α_k\} uses these reference layers, not current model predictions; the residual target is R = I_src − I_c.

5.2 Training, inference, and the framework diagram

Training inputs are clean I_src, noisy R R=Isrc−IcR = I_src − I_c, noisy B/FG layers, and noisy degraded FGs. The regular path generates layers from noise; the auxiliary path restores clean FGs from noisy degraded FGs. Each degraded FG shares its target FG’s positional encoding and attends only to itself and I_src. Inference removes this auxiliary path and generates BG, FGs, and R from I_src. We will diagram the training inputs/targets and inference process.

6. What do Layers 1 and 2 mean in Figure 5? How are the attention weights obtained? Are they averaged?

6.1 What do Layer 1 and Layer 2 represent?

Layers 1 and 2 are the first two FG images, represented by two token groups in the noisy layers.

6.2 How are attention weights extracted and averaged?

Figure 5 uses the penultimate DiT block: noised I_src tokens query Layer 1, Layer 2, and clean I_src. We average attention over each target image’s key tokens to compare trends. We will document extraction/aggregation and clarify labels.

7. Writing: identify the backbone earlier and clarify the method diagram and figure labels

We apologize for unclear framework/figure descriptions. Revised Section 4 will identify Qwen-Image-Layered [3] as the backbone, diagram the method/inputs, unify text/figure notation, and clarify module roles and training–inference differences.

References

[1] Shuai Bai et al.. Qwen3-VL Technical Report. arXiv:2511.21631, 2025.

[2] Maxime Oquab et al.. DINOv2: Learning Robust Visual Features without Supervision. Transactions on Machine Learning Research, 2024.

[3] Shengming Yin et al.. Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition. arXiv:2512.15603, 2025. Official model weights.

Reviewer_ddviOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses layer decomposition for in-the-wild natural images. The authors first introduce an automated Agent-driven Data Decomposition (ADD) pipeline that chains off-the-shelf foundation models (Qwen3-VL-32B, FLUX.2-klein-9B, SAM3, and RMBG-1.4/2.0) with DINOv2-based heuristic filtering and VLM verification to curate LiWi-100k, a dataset of over 100,000 layered images. Second, they fine-tune Qwen-Image-Layered on LiWi-100k with two methodological modifications: replacing the source-image reconstruction target with a pixel-space photometric residual layer (R=Isrc−IcR = I_{src} - I_c), and introducing an auxiliary degradation-restoration objective that recovers clean foregrounds from boundary-corrupted inputs during training. Experiments are conducted on the LiWi-100k test split, the graphic-design benchmark Crello, and the DIS-5K segmentation dataset.

Critical strengths
  • Scalable data pipeline and dataset release: Decoupling background/foreground generation with DINOv2 filtering and VLM verification in the ADD pipeline is a pragmatic engineering effort for curating natural-image RGBA supervision. The promised release of LiWi-100k and code will benefit the community.

  • Identifying the reconstruction shortcut in Qwen-Image-Layered: The authors empirically show (Table 4, Fig. 5) that jointly reconstructing the source image IsrcI_{src} induces an attention shortcut (information leakage); removing this objective substantially improves layer disentanglement.

Critical weaknesses

1.Benchmark Limitations & Evaluation Protocol Flaws

  • 1.1 Lack of out-of-distribution real-world benchmarks and unfair Crello comparisons: First, the LiWi-100k test split shares the same synthetic ADD pipeline as the training data without human-verified or physically captured GT, introducing pipeline bias. Meanwhile, Crello (graphic designs, Table 2) and DIS-5K (binary foreground segmentation without background inpainting or amodal completion, Table 3) do not evaluate holistic natural-image decomposition, leaving out independent real/semi-real layered benchmarks (e.g., LayerDecomp, MuLAn, or object removal datasets). Second, evaluating on Crello contradicts the paper's in-the-wild motivation: as noted in Section 1, graphic designs feature "clean boundaries, explicit layer ordering, and simple alpha blending," where R=Isrc−Ic≡0R = I_{src} - I_c \equiv 0 largely holds, invalidating the photometric residual premise. Moreover, Table 2 is unfair because LiWi-Crello is fine-tuned on Crello for 6K extra steps (Section 5.1) while Qwen-Image-Layered is not; controlled baselines fine-tuning both Qwen-Image-Layered and Qwen-Image-Layered-SFT on Crello for 6K steps are missing.

  • 1.2 Conflated foreground/background evaluation: Unlike standard protocols (e.g., LayerDecomp) that evaluate foreground (FG) and background (BG) layers separately, Table 1 averages FG and BG into a single sequence score (Appendix B, Eq. 8), obscuring individual layer quality. Separate FG and BG evaluations on public benchmarks are needed.

  • 1.3 Deficiencies in evaluation metrics:

  • a. Unseparated occluded vs. visible regions: Averaging RGB L1 over all nonzero-alpha pixels (Appendix B, Eq. 4) covers the entire image for an opaque background (αB≡1\alpha_B \equiv 1), allowing unoccluded pixels to dilute errors in small disoccluded or shadow/reflection regions. Region-specific metrics (e.g., Masked/Occluded RGB L1, LPIPS, FID) should be reported.

  • b. Inadequacy of pixel-wise L1 for generative completion: Amodal foreground and background completion is an ill-posed generative task; relying solely on RGB L1 penalizes plausible high-frequency textures that deviate from the pseudo-GT while favoring blurry predictions.

  • c. Missing recomposition and perceptual evaluations: The paper reports neither inference-time recomposition error (comparing recomposed layers against IsrcI_{src}, with or without R^\hat{R}) nor human/VLM preference studies.

  • 1.4 Selective metric reporting and missing ablations on DIS-5K: Table 3 only reports region-dominated metrics that are insensitive to fine boundaries, relegating the crucial boundary metric HCEγHCE_\gamma to Appendix J (Table 11), where LiWi (1359 on TE(1-4)) trails BiRefNet (916) and even 2020 baselines like BASNet (1007) and U^2-Net (1042). Furthermore, while the Degradation-Restoration Objective targets boundary defects (erosion/dilation), its ablation (Table 4) is restricted to synthetic LiWi-100k using global Alpha soft IoU. The authors should provide the Table 4 ablations (including Qwen-Image-Layered-SFT and LiWi w/o Degradation-Restoration) on DIS-5K with HCEγHCE_\gamma to verify actual boundary gains on real GT masks.

2.Limited Algorithmic Contribution & Overstated Novelty

  • Marginal algorithmic gains and overstated novelty: In Tables 1 and 4 (#Max Edits = 0), fine-tuning on LiWi-100k (Qwen-Image-Layered →\rightarrow Qwen-Image-Layered-SFT) drives the vast majority of the gain, reducing RGB L1 by 63.00% (0.2692 →\rightarrow 0.0996) and boosting Alpha soft IoU by 48.78% (0.5878 →\rightarrow 0.8745). In contrast, the proposed modeling changes (Qwen-Image-Layered-SFT →\rightarrow LiWi) add only 8.94% (0.0996 →\rightarrow 0.0907) and 6.51% (0.8745 →\rightarrow 0.9314) relative gains, most of which simply come from dropping - Source Image Reconstruction (0.0950 RGB L1, 0.9177 Alpha soft IoU). Adding + Degradation-restoration Objective improves RGB L1 by a negligible 0.0003 (0.0910 →\rightarrow 0.0907), which likely falls within random seed variance and fails to support the claimed color/texture refinement. Moreover, synthesizing xdx_d via boundary erosion, dilation, and blurring is a standard morphological data augmentation rather than a novel formulation.

3.Insufficient Dataset Validation & Scalability Analysis

  • 3.1 Insufficient comparison with public layered datasets (e.g., MuLAn): Since LiWi-100k drives most of the performance gain, it should be benchmarked against existing open-source natural-image layered datasets like MuLAn. Instead of only auditing a re-implemented MuLAn pipeline without its failure classifiers (Appendix F, Table 8), the authors should train identical backbones on the official MuLAn dataset vs. LiWi-100k and compare them on third-party benchmarks.

  • 3.2 Lack of data-scaling analysis: Despite claiming high scalability, the paper only evaluates a single data scale (100k) without scaling ablations (e.g., 10k, 25k, 50k, 100k, 200k) to show whether performance has already saturated under the synthetic pseudo-GT ceiling of the teacher models (FLUX.2-klein-9B and RMBG/SAM3).

Questions for authors

Please take a look at the weaknesses. That’s all I wanted to ask.

4 comments
AuthorsOpenReview ↗

Part 1: W1.1

Dear Reviewer,

Thank you for your detailed review. We added test-set human verification, external MuLAn evaluation, Crello comparisons with matched adaptation steps, and additional ablations to strengthen the evaluation.

Benchmark Limitations & Evaluation Protocol Flaws

1.1 Lack of out-of-distribution real-world benchmarks and unfair Crello comparisons

1.1.1 Human verification of the test-set reference layers

The original manuscript lacks independent verification of the full test set. Three annotators independently assessed all 1,000 images in test set for FG completeness, BG completion, alpha boundaries, and recomposition consistency. We will manually revise failed reference layers. Appendix A’s 94.9% average human pass rate supports overall data quality.

The results are shown in the table below.

AssessmentForeground completenessBackground cleanup and completionAlpha boundariesRecomposition consistencyOverall pass
Annotator 1972/1,000(97.2%)978/1,000(97.8%)965/1,000(96.5%)973/1,000(97.3%)951/1,000(95.1%)
Annotator 2973/1,000(97.3%)980/1,000(98.0%)976/1,000(97.6%)972/1,000(97.2%)958/1,000(95.8%)
Annotator 3971/1,000(97.1%)970/1,000(97.0%)971/1,000(97.1%)969/1,000(96.9%)951/1,000(95.1%)
Majority vote977/1,000(97.7%)983/1,000(98.3%)975/1,000(97.5%)971/1,000(97.1%)950/1,000(95.0%)

1.1.2 Independent natural-image layered benchmark: MuLAn

MuLAn [1] does not explicitly handle object-removal effects on BG eg˙,˙shadows,asnotedinitspapere\.g\., shadows, as noted in its paper; LiWi construction/verification removes object-related shadows/reflections. We have added external evaluation on MuLAn.

The table below shows held-out MuLAn FG extraction under zero-shot evaluation and equal-budget adaptation.

ModelMuLAn additional training stepsPSNR ↑SSIM ↑LPIPS [2] ↓FID [3] ↓
Qwen-Image-Layered025.640.9470.10459.28
Qwen-Image-Layered-SFT027.380.9600.08447.83
LiWi027.910.9660.07541.62
Qwen-Image-Layered → MuLAn6K27.960.9670.07340.85
Qwen-Image-Layered-SFT → MuLAn6K28.570.9720.06536.47
LiWi → MuLAn6K29.060.9760.05832.81

BG removal/completion results under the same settings are below.

ModelMuLAn additional training stepsPSNR ↑SSIM ↑LPIPS [2] ↓FID [3] ↓
Qwen-Image-Layered022.830.8170.18165.24
Qwen-Image-Layered-SFT024.620.8720.13949.76
LiWi025.180.8890.12444.38
Qwen-Image-Layered → MuLAn6K25.730.9040.11239.65
Qwen-Image-Layered-SFT → MuLAn6K26.490.9180.09632.87
LiWi → MuLAn6k27.08<br>0.9290.08428.43

1.1.3 Does Crello evaluation conflict with the natural-image motivation?

Crello [4] tests transfer to graphic design. R is the source–layer-composition difference: it can reflect shadows/illumination in natural images or boundaries/antialiasing in designs, and is zero when source and composition match. Even when R is near zero in Crello, the residual branch can still act as an attention sink, absorbing attention associated with regions of I_src that are irrelevant to individual layers[5]. In Table 2, the fully adapted system averages 12.45% lower RGB L1 and 1.35% higher alpha soft IoU than the listed Qwen baseline, supporting generality.

AuthorsOpenReview ↗

Part 2: W1.1-W1.3

1.1.4 Crello adaptation for 6K steps and controlled baselines

Our Table 2 quotes Qwen-Image-Layered’s Table 1 [6]; its Section 4.3.1 specifies Crello fine-tuning, so this is not a zero-shot baseline. For a reproducible equal-budget comparison, we fine-tune public Qwen-Image-Layered weights on Crello for 6K additional steps and evaluate Max Edits = 0–5 under the same protocol. Below we compare the reported baseline, public-weights 6K control, and LiWi-Crello.

RGB L1 results are below.

Model / result sourceCrello stepsEdits 0Edits 1Edits 2Edits 3Edits 4Edits 5
Qwen-Image-Layered-I2LUnknown0.05940.04900.03930.03770.03640.0363
Public Qwen-Image-Layered → Crello6K0.06060.05000.04020.03850.03710.0369
LiWi-Crello6K0.05120.04230.03450.03310.03230.0321

Alpha soft IoU results are below.

Model / result sourceCrello stepsEdits 0Edits 1Edits 2Edits 3Edits 4Edits 5
Qwen-Image-Layered-I2LUnknown0.87050.88630.91050.91210.91560.9160
Public Qwen-Image-Layered → Crello6K0.86660.88280.90710.90900.91270.9132
LiWi-Crello6K0.88230.89560.92240.92500.92790.9308

1.2 Conflated foreground/background evaluation

We follow LayerD [7] (Appendix B):

Max Edits = 0: group layers by visibility, match predicted/reference layers, and compute RGB L1 per pair, including FG/BG, without additional adjacent-layer merging.

Max Edits > 0: allow up to that many additional adjacent-layer merges after grouping/matching eg˙,˙atmostoneforMaxEdits=1e\.g\., at most one for Max Edits = 1. Evaluate local recompositions alongside unmerged layers.

Evaluation thus includes FG, BG, and possible partial layer compositions.

1.3 Deficiencies in evaluation metrics:

1.3.1 Metrics for visible regions, occluded regions requiring completion, and shadow/reflection regions

Large visible regions can dominate overall RGB error. The table below separately reports BG-visible, BG-completion, and FG-completion RGB L1, using identical reference masks across methods and normalizing by valid pixels in each region.

MethodBG visible L1 ↓BG completion L1 ↓FG completion L1 ↓
Qwen-SFT0.0360.1860.158
Remove source-image reconstruction0.0370.1790.152
Add pixel-space residual0.0390.1690.148
Full LiWi0.0390.1700.147

BG visible: background visible in I_src. BG completion: background hidden by FG objects and completed after removal. FG completion: parts of a FG object hidden by other objects and requiring completion.

Full LiWi versus Qwen-SFT lowers BG-completion L1 from 0.186 to 0.170 and FG-completion L1 from 0.158 to 0.147, while BG-visible L1 rises slightly from 0.036 to 0.039.

1.3.2 Does pixel-wise L1 penalize plausible alternative completions and favor blurry results?

Results cover all images. DWT+L1 combines L1 only over high-frequency subbands LH, HL, and HH; LiWi improves the measured metrics.

High-frequency results are below.

ModelGMSD [8] ↓DWT+L1 ↓
Qwen-Image-Layered0.23340.0198
LiWi0.12770.0118

1.3.3 Inference-time recomposition with and without the predicted shadow

We alpha-composite predicted FGs with predicted BG and measure RGB L1, PSNR, and LPIPS against I_src. The table below compares Qwen-SFT and three LiWi variants, all using RGBA composition at inference.

Method / recompositionRGB L1 ↓PSNR ↑LPIPS ↓
Qwen-SFT / RGBA0.014833.60.0526
LiWi without reconstruction / RGBA0.012734.90.0448
LiWi + predicted shadow / RGBA only0.011535.70.0372
Full LiWi / RGBA0.010336.50.0308
AuthorsOpenReview ↗

Part3: W1.4-W3

1.4 Selective metric reporting and missing ablations on DIS-5K

Space limits moved complete metrics, including HCEγ, to Appendix J, Table 11. We will add HCEγ to main-paper Table 3 and DIS-5K [9] comparisons with Qwen-Image-Layered-SFT and LiWi w/o Degradation-Restoration to assess restoration’s effect on real-image segmentation.

DIS-5K metrics and degradation-restoration ablations are below.

SplitModelFβˣ ↑Fβʷ ↑M ↓Sₘ ↑Eφᵐ ↑HCEγ ↓
TE 1–41–4Qwen-Image-Layered-SFT0.7690.7150.0840.8020.8501518
TE 1–41–4LiWi w/o Degradation-Restoration0.7870.7350.0850.8150.8641542
TE 1–41–4LiWi (original paper)0.7940.7440.0770.8210.8711359
VDQwen-Image-Layered-SFT0.7660.7140.0860.8010.8491586
VDLiWi w/o Degradation-Restoration0.7840.7340.0870.7990.8631498
VDLiWi (original paper)0.7910.7440.0780.8210.8711417

2. Limited Algorithmic Contribution & Overstated Novelty

SFT’s substantial gains demonstrate the data’s value. With identical data, full LiWi further averages 7.99% lower RGB L1 and 6.22% higher alpha soft IoU than SFT across Max Edits = 0–5. Gains thus also reflect the training design: identifying/mitigating the source-reconstruction shortcut, Shadow-Guided Learning, and boundary refinement through degradation-restoration.

3. Insufficient Dataset Validation & Scalability Analysis

3.1 Insufficient Dataset Validation & Scalability Analysis

See 1.1.2.

3.2 Lack of data-scaling analysis

We trained on 10K, 25K, 50K, and 100K samples for 12K optimization steps each. LiWi test results are below.

RGB L1 by data scale at Max Edits = 0–5 is below.

Data scaleStepsEdit 0Edit 1Edit 2Edit 3Edit 4Edit 5
10K12K0.11100.11050.10890.10780.10690.1062
25K12K0.10100.10070.09940.09850.09780.0973
50K12K0.09400.09380.09260.09190.09130.0909
100k12K0.09070.09060.08950.08890.08830.0879

Alpha soft IoU under the same settings is below.

Data scaleStepsEdit 0Edit 1Edit 2Edit 3Edit 4Edit 5
10K12K0.88900.89010.89240.89400.89520.8961
25K12K0.91200.91280.91460.91580.91680.9175
50K12K0.92400.92460.92610.92710.92790.9286
100k12K0.93140.93190.93320.93400.93470.9353

More data improves both metrics across Max Edits = 0–5. At Max Edits = 0, RGB L1 falls from 0.1110 to 0.0907 and alpha soft IoU rises from 0.8890 to 0.9314. The training pool is limited to 100K, so 200K was not evaluated.

References

[1] Tudosiu et al. MULAN: A Multi Layer Annotated Dataset for Controllable Text-to-Image Generation. CVPR, 2024.

[2] Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR, 2018.

[3] Heusel et al. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. NeurIPS, 2017.

[4] Yamaguchi. Crello dataset. Hugging Face, accessed 2026.

[5] Gu et al. When Attention Sink Emerges in Language Models: An Empirical View. ICLR, 2025.

[6] Yin et al. Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition. arXiv, 2025.

[7] Suzuki et al. LayerD: Decomposing Raster Graphic Designs into Layers. ICCV, 2025.

[8] Xue et al. Gradient Magnitude Similarity Deviation: A Highly Efficient Perceptual Image Quality Index. IEEE TIP, 2014.

[9] Qin et al. Highly Accurate Dichotomous Image Segmentation. ECCV, 2022.

Thank you for the additional experiments. The MuLAn evaluation, DIS-5K ablation, and data-scaling results are helpful. I acknowledge that the original Qwen baseline was fine-tuned on Crello. The following concerns remain:

  1. Controlled comparisons. The MuLAn experiments do not isolate the value of LiWi-100k versus official MuLAn training data under matched initialization, objectives, and training budgets on a common independent benchmark. Crello also lacks Qwen-Image-Layered-SFT → Crello with the same adaptation budget as LiWi-Crello, needed to distinguish data gains from method gains.

  2. Recomposition versus correct separation. Section 1.3.3 describes results “with and without” the predicted shadow, but the table appears to report RGBA-only composition. Please compare composition with versus without the predicted residual using the same model outputs. Crucially, good recomposition does not establish correct shadow separation: shadows left in the background can still yield an accurate recomposed image. Separate shadow/reflection-region evaluation and visualizations of the cleaned background and residual are therefore needed.

  3. Completion quality. The MuLAn LPIPS/FID results are welcome, but GMSD and DWT+L1 still penalize deviations from a particular reference and do not fully assess plausible alternative completions. The high-frequency comparison also lacks Qwen-SFT. Please include a focused perceptual comparison of completion regions against this matched baseline, ideally with human preference evaluation.

  4. Boundary and transparency evidence. The reported DIS-5K HCEγ reduction from 1542 to 1359 supports the restoration objective, although LiWi still trails specialist segmentation baselines. Given the emphasis on alpha-boundary accuracy, could you show challenging cases involving hair/fur, bicycle spokes, tennis-racket strings, and semi-transparent drinking glasses? Enlarged crops, alpha maps, and comparisons with Qwen-SFT and LiWi without degradation-restoration would make the improvements easier to assess.

  5. Evaluation details. Were the reported results computed against the original references or the manually corrected ones? All methods should use the same finalized references. Please also specify the MuLAn split, evaluation sample count, and FG metric computation, particularly the treatment of transparent regions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.