acceptodds
Under review as a conference paper at ICLR 2027

Editing Images into Reusable Layers via Compositional Edit-Prior Distillation

Abstract

High-quality image editing requires precise modification of target regions while preserving the content of unrelated regions. Existing editing models produce flattened outputs that limit precise local manipulation and layer reuse, while decomposition methods typically optimize layer recovery separately from instruction-driven editing. To this end, we introduce , a unified framework that couples instruction-guided image editing with controllable generation of reusable RGBA layers. At its core, jointly represents the edited canvas and an ordered sequence of RGB–Alpha layers within a single token sequence, enabling semantic edits and explicit control over individual layer content. To retain the pretrained model's editing capability during adaptation to layered outputs, we propose , which transfers its editing prior to both the edited canvas and the image composed from the predicted layers. To support training and evaluation, we construct , a large-scale paired image-editing dataset with ordered RGBA layer annotations, and introduce dedicated benchmarks for layered editing. Extensive experiments demonstrate that LayerEditor maintains strong editing fidelity while generating accurate RGBA layers, supporting both flexible layer-level manipulation and multi-layer decomposition.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Editing, Layered Image Generation, Edit-Prior Distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nKreOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces LayerEditor, a framework that jointly performs instruction-guided editing of a source image and generates an ordered stack of RGBA layers. It extends a pretrained image editor through joint RGB–Alpha token modeling and a separately trained Alpha decoder. Its Compositional Edit-Prior Distillation (C-EPD) queries a frozen editing teacher on student-visited states, supervising both the student’s direct canvas velocity and the decoded composition of its predicted layers. The authors construct LayerEditor-100K and evaluate on two layered-editing benchmarks, layer-count and instruction-control settings, and image-to-layer decomposition. The reported results show improved recomposition fidelity and layer quality over the evaluated baselines, with ablations supporting the complete C-EPD configuration.

Critical strengths
  1. A useful framework for jointly editing images and generating controllable layers. This paper proposes LayerEditor, a system that generates an edited target image together with its ordered RGBA layer decomposition, supporting user-specified layer counts and layer-content descriptions. This structured output makes individual components accessible for subsequent manipulation, addressing a practical limitation of conventional editors that produce only flattened images.

  2. A substantial paired dataset for instruction-guided layered editing. This paper introduces LayerEditor-100K, a dataset that covers text, object, and appearance edits with 2–10 layers per example, providing supervision for both editing outcomes and their underlying layer structures. The accompanying benchmarks evaluate recomposition fidelity, layer quality, and controllability. This data contribution is consequential because it addresses the scarcity of paired editing-and-layer annotations and provides a concrete foundation for training and evaluating models for this task.

Critical weaknesses
  1. The central claim of reusable layers is not directly validated in downstream manipulation. The quantitative evaluation primarily measures reconstruction, reference-layer agreement, and VLM-assessed quality in the original layouts. It does not systematically test whether moving, removing, resizing, or replacing generated layers preserves usable boundaries, recovers newly exposed content, and avoids residual artifacts.
  2. Limited validation of instruction-driven semantic grouping. Although LayerEditor supports layer-content descriptions and reports semantic-grouping scores in Table 2, the results do not clearly establish whether it can flexibly group multiple instances of the same category according to user intent. For example, can two birds be placed together in one layer when requested, or separated into different layers under an alternative instruction? A controlled comparison that varies the requested grouping while keeping the source image and intended edited appearance fixed would help distinguish instruction-driven grouping from category-based or instance-based decomposition.
  3. Limited validation of editing-capability retention. The five rubric dimensions are scored by a single Gemini-3.1-Pro evaluator without validation against human judgments (Appendix C.3). Figure 11 illustrates score anchors but does not establish agreement with human assessments. Although reference-based metrics provide complementary evidence, human blind evaluation and metric computation on an independent editing benchmark would strengthen conclusions about perceptual quality and editing-capability retention.
Questions for authors

Please review the weaknesses noted in my comments. I am open to increasing my score if the authors can properly address these concerns.

4 comments
AuthorsOpenReview ↗

1. Direct evaluation of downstream layer reuse (W1)

We thank the reviewer for the constructive suggestion. We conducted a method-blind human evaluation involving 50 evaluators with professional design backgrounds, each reviewing the same 500 cases randomly sampled from Crello-Testset and ScaleEdit-Testset. Each case includes at least three successive rounds of layer reuse, with edits carried forward between rounds. Operations include insertion, deletion, transformation, reordering, and text-layer editing. All methods underwent the same sequences with the same parameters. The modified stacks are recomposed without regeneration or repair, allowing evaluators to inspect layer boundaries, residual fragments, and newly exposed content (Section 4.6, p. 10; Appendix C.4, p. 23).

As shown in Table 5 (p. 10), LayerEditor achieves 7.97 for editing reusability, compared with 6.13 without C-EPD and 4.81 for the strongest external baseline. We found that unintended content mixing and blurred layer boundaries commonly hinder competing methods during subsequent reuse. Figure 8 (p. 10) shows independent movement and resizing of the turkey and basket. Figures 15–16 (pp. 27–28) show different requested layer organizations and successive manipulation of the resulting layers.

Table 5. Human evaluation of layered editing (p. 10). Scores range from 1 to 10, with higher values indicating better results. Edit., VQ., Cont., Comp., and Reuse. denote edit consistency, visual quality, layer-content accuracy, layer-composition accuracy, and editing reusability, respectively.

MethodEdit. ↑VQ. ↑Cont. ↑Comp. ↑Reuse. ↑
Seedream5.975.575.605.174.80
FR + QIL5.636.072.605.573.50
FR + QIL-Decomp-FT3.973.274.204.233.88
FR + QIL-Control4.124.254.233.874.81
LayerEditor w/o C-EPD5.836.436.176.536.13
LayerEditor7.177.337.677.977.97
AuthorsOpenReview ↗

2. Instruction-driven semantic grouping (W2)

LayerEditor uses text instructions to specify layer count and content. The Target layers section describes each layer in back-to-front order, with one entry per predicted RGB–Alpha pair. For example, L1: both people assigns both people to a single layer so they can be moved or scaled together. In contrast, L1: left person and L2: right person request separate layers, allowing either person to be moved, resized, or removed independently while leaving the other person unchanged. The layer count and descriptions must agree. Training and inference use the same four-section prompt, with an optional Prompt Enhancer assisting users who provide only a brief editing request (Appendix A.1, p. 14). Figures 15–16 (pp. 27–28) illustrate language-controlled grouping of people, buildings, and food items, followed by joint or independent manipulation. With the source image and intended edited appearance held fixed, LayerEditor achieves a semantic-grouping score of 6.88, compared with 5.91 for Seedream5-Pro (Joint) (Table 2, p. 8).

Table 2. Layer-count and semantic-grouping control with the source image and intended edited appearance held fixed (main paper, p. 8). Sem., Comp., and Alpha denote semantic-grouping, layer-composition, and Alpha-accuracy rubric scores on a 1–10 scale.

MethodRGBA L1 ↓Alpha IoU ↑Sem. ↑Comp. ↑Alpha ↑
Seedream5-Pro (Seq.)0.11840.49845.524.682.77
Seedream5-Pro (Joint)0.10280.57035.915.154.28
FR + QIL0.16250.23784.303.662.02
FR + QIL-Control0.11750.42855.474.562.56
FR + QIL-Decomp-FT0.19870.25545.655.471.36
QIL-Edit-SFT0.11480.30965.104.571.60
RevealLayer0.40960.22083.213.751.11
LayerEditor0.04600.61966.886.286.98

Appendix D.1 (Table 8, p. 24) further varies editing instructions across 3,994 source documents, with a corresponding target image and reference RGBA stack for each request. Scores are averaged within documents before aggregation. Compared with Seedream5-Pro (Joint), LayerEditor improves edit consistency by 66.9% (3.32 → 5.54) and Composite SSIM by 10.8% (0.7914 → 0.8765), while reducing RGBA L1 by 38.9% (0.0563 → 0.0344). These improvements support LayerEditor’s ability to follow revised editing instructions while generating the corresponding layer organization.

Table 8. Instruction controllability across 3,994 source documents from Crello-Testset and ScaleEdit-Testset (Appendix D.1, p. 24). Editing instructions vary for each source image, and each prediction is evaluated against the corresponding target image and reference RGBA stack.

MethodComposite SSIM ↑RGBA L1 ↓Alpha IoU ↑Edit. ↑Seg. ↑Comp. ↑
Seedream5-Pro (Joint)0.79140.05630.56463.324.813.74
FR + QIL0.47770.14350.19451.163.962.67
FR + QIL-Control0.58270.09840.35021.455.113.87
FR + QIL-Decomp-FT0.59720.17330.21231.414.684.34
QIL-Edit-SFT0.70460.19780.21041.223.633.32
LayerEditor0.87650.03440.58235.546.335.68
AuthorsOpenReview ↗

3. Human Evaluation of Editing Capability and Layer Reusability (W3)

Thank you for the constructive suggestion. ScaleEdit-Testset is built specifically from editing image pairs, with 1,000 held-out source documents and corresponding target layers (Section 3.4, p. 6; Appendix B.1–B.3, pp. 17–18). ScaleEdit-Testset tests whether requested edits remain correct after layer recomposition. In Table 1 (p. 6), LayerEditor improves Composite SSIM by 5.7% and reduces Composite FID by 33.8% relative to Seedream5-Pro (Joint). LayerEditor also achieves the highest rubric-based edit-consistency score, while the matched ablations confirm the editing improvements from C-EPD (Table 3, p. 9).

We added a method-blind human evaluation in which 50 evaluators with professional design backgrounds reviewed the same 500 cases sampled from both benchmarks. Evaluators scored editing, visual quality, layer content, composition, and reuse on a 1–10 scale (Section 4.6, p. 10; Appendix C.4, p. 23). LayerEditor leads all five dimensions in Table 5 (p. 10, shown above). Compared with w/o C-EPD, the mean human scores increase by 23.0% for edit consistency and 14.0% for visual quality, supporting the improvement trends observed with Gemini. Figures 11–12 (pp. 22–23) provide the Gemini rubric and scoring examples.

The user study also evaluates at least three consecutive rounds of layer reuse, with identical operation sequences and parameters across methods and no regeneration or repair. LayerEditor’s mean reuse score is 7.97, approximately 30.0% higher than w/o C-EPD and 65.7% higher than the strongest external baseline, FR + QIL-Control. Figure 8 (p. 10) illustrates independent layer operations, and Figures 15–16 (pp. 27–28) show different layer organizations and multi-round reuse. These results support both faithful initial editing and continued control of the generated layers, with usable recompositions after subsequent manipulation.

Table 1, ScaleEdit-Testset metrics (p. 6).

MethodComposite SSIM ↑Composite FID ↓Edit Cons. ↑Visual Qual. ↑
Seedream5-Pro (Seq.)0.790170.694.654.80
Seedream5-Pro (Joint)0.836559.624.054.71
FR + QIL0.645181.094.924.76
FR + QIL-Control0.660684.074.705.08
RevealLayer0.646475.744.284.08
LayerEditor0.883939.485.966.58

Table 3. Ablations on Crello-Testset (p. 9).

ModelEdit Cons. ↑Visual Qual. ↑Layer Seg. ↑RGBA L1 ↓
LayerEditor6.146.356.830.0350
w/o C-EPD4.675.125.010.0482
w/o Direct OPD5.014.194.830.0513
w/o Comp. OPD (Vanilla on-policy OPD)4.984.334.970.0492
w/o Direct latent4.354.654.960.0642
Base Model3.934.133.770.0735

Dear Reviewer nKre,

Thank you very much for your time and valuable feedback.

We understand that you may be busy. With the discussion deadline approaching and only one day remaining, we would appreciate it if you could review our responses and the revised manuscript.

Please let us know if any of your concerns remain unresolved. We would be very happy to address any further questions you may have.

Thank you for your attention.

Best regards, Authors of #7343

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

This paper proposes LayerEditor, an instruction-guided image editing framework that generates multiple composable RGBA layers that are composed as the edited image target, instead of generating a single target image alone. The framework is finetuned from a pretrained image editing framework, with a modified input token sequence consisting of the direct edited canvas and each layer's RGB and Alpha mask. In order to retain the pretrained model's knowledge on image editing after the token sequence modification, the paper introduces a compositional edit-prior distillation (C-EPD) that aligns both the direct canvas' prediction and the composite layer's prediction with the pretrained model (teacher)'s target. Results show that the C-EPD has effectively reduced edit inconsistency.

Critical strengths

The proposed model architecture is reasonable, which modifies the model's input and output token sequence with clear role assignments for each component. The main issue with this modification, which is the ability to retain the pretrained model's image-editing knowledge, is well addressed by the proposed C-EPD on-policy distillation approach. The ablation study has validated this. The results look promising. The task is meaningful due to the interpretability of each layer and the flexibility for layer reuse.

Critical weaknesses
  1. The model requires a pre-defined number of layers as input. How to decompose the image into layers is an ill-posed problem, while using a pre-defined layer number may result in a bad layer decomposition. See Question 1 for examples.
  2. It is unclear what the role of the direct edit canvas is in the output sequence, with the presence of the composite image from the layer stack being the final edited image. No results compare the composite image and the direct edit canvas.
Questions for authors
  1. How does the model automatically decide how to assign the role of each layer, given a pre-defined fixed layer number?
    • Example 1: In the teaser (Fig. 1), the curved underline below "Venice" is assigned to the "Decorations" layer rather than "Text", but assigning it to "Text" is also reasonable. In addition, how will the model output the layers if we give a different layer number 5 to the same image? Will the layer decomposition given layer number 5 be consistent with the decomposition given 6 by merging two layers in it, or will it give a completely different layer decomposition?
    • Example 2: In the lower row of Fig. 5 (the bird), it is an ill-posed problem to decide how the tree branches, flowers and leaves should be assigned to each layer.
  2. What is the role of the direct edit canvas? Is it purely used for the Direct OPD part within C-EPD? Is it discarded during inference time? Can the authors provide a comparison between the direct edit canvas output and the composed layers to see if they're consistent?
  3. In all the results, the Alpha masks appear to be clean with sharp and consistent borders. How does the method guarantee this? I assume that the pretrained model should provide no knowledge on generating clean Alpha masks.
5 comments
AuthorsOpenReview ↗

1. Prompt-driven layer organization and flexible layer counts (W1, Q1)

LayerEditor uses instructions to specify both the layer count and the content of each layer. The count controls granularity, while ordered descriptions determine which elements are grouped or separated. For example, L1: both people assigns both people to a single layer so they can be moved or scaled together. In contrast, L1: left person and L2: right person request separate layers, allowing either person to be moved, resized, or removed independently while leaving the other person unchanged. The count and descriptions must agree, as conflicting specifications may cause unintended grouping. Training and inference use four prompt sections, Instruction, Preserve, Target layers, and Output. For brief requests without layer descriptions, the optional Prompt Enhancer helps construct this structure (Appendix A.1, p. 14).

For the Venice example (Figure 1, p. 2), “Venice lettering and the curved underline” groups both in the text layer, while “curved underline and other ornaments” assigns the underline to decoration. To request five rather than six layers, the user combines two layer descriptions and retains the other four. LayerEditor jointly generates the five specified layers from the source image, rather than merging previous predictions. The same text-based control specifies how branches, flowers, and leaves are grouped in Figure 5 (p. 8). We additionally show six-layer and four-layer decompositions of the bird illustration in Figure 15, extending the second example in Figure 5 with different text-specified groupings and subsequent multi-round reuse. Table 2 (p. 8) evaluates layer organization while holding the source image and intended edited appearance fixed. LayerEditor achieves a semantic-grouping score of 6.88, compared with 5.91 for the strongest baseline.

Table 2. Layer-count and organization control (p. 8). Sem., Comp., and Alpha are 1–10 rubric scores for semantic grouping, layer composition, and Alpha accuracy.

MethodRGBA L1 ↓Alpha IoU ↑Sem. ↑Comp. ↑Alpha ↑
Seedream5-Pro (Seq.)0.11840.49845.524.682.77
Seedream5-Pro (Joint)0.10280.57035.915.154.28
FR + QIL0.16250.23784.303.662.02
QIL-Edit-SFT0.11480.30965.104.571.60
LayerEditor0.04600.61966.886.286.98

Appendix D.1 (Table 8, p. 24) additionally varies editing instructions across 3,994 source documents, with a corresponding target image and reference RGBA stack for each request. Scores are averaged within each document before aggregation. LayerEditor achieves 5.54 for edit consistency versus 3.32 for Seedream5-Pro (Joint), and 6.33 for segmentation versus 5.11 for FR + QIL-Control. Together, these experiments support control over what is edited and how the result is organized into layers, with reference-based reconstruction metrics complementing the rubric scores. Figures 15–16 (pp. 27–28) illustrate different requested layer counts and semantic groupings, followed by multi-round reuse.

Table 8. Instruction controllability across 3,994 source documents (Appendix D.1, p. 24). Composite SSIM measures similarity to the edited target. Edit., Seg., and Comp. are 1–10 rubric scores for edit consistency, layer segmentation, and layer composition.

MethodComposite SSIM ↑RGBA L1 ↓Alpha IoU ↑Edit. ↑Seg. ↑Comp. ↑
Seedream5-Pro (Joint)0.79140.05630.56463.324.813.74
FR + QIL0.47770.14350.19451.163.962.67
QIL-Edit-SFT0.70460.19780.21041.223.633.32
LayerEditor0.87650.03440.58235.546.335.68
AuthorsOpenReview ↗

2. Roles of the Direct canvas and comparison with the Composite output (W2, Q2)

The Direct canvas serves as a global attention sink and semantic anchor, as well as the teacher interface for C-EPD. Through bidirectional attention, the canvas and individual layers exchange whole-image context, including layout, object relationships, and editing intent, to coordinate layer generation. During C-EPD, the frozen pretrained editor receives the student’s Direct latent and predicts a velocity that supervises the Direct branch. The same teacher prediction is used to estimate and decode a clean-image target for the RGBA composition, extending image-level editing guidance to the layered output (Sections 3.1–3.2, pp. 3–4).

At inference, the teacher is removed, while the Direct latent remains in joint generation to preserve the learned attention context. The final Composite is rendered solely from the RGBA layers. Table 1 (p. 6) compares Direct and Composite against the same edited target. Direct and Composite achieve SSIM scores of 0.8530 and 0.8351 on Crello-Testset, respectively, and 0.8925 and 0.8839 on ScaleEdit-Testset, indicating that layer recomposition largely preserves target fidelity. Removing Direct only at inference reduces Alpha accuracy from 5.05 to 4.25 (15.8%) and layer-composition quality from 5.64 to 4.17 (26.1%). These declines support retaining the Direct latent to maintain cross-layer coordination during inference.

We further evaluate a w/o Direct latent variant that removes the Direct output slot during both training and inference while retaining RGB–Alpha flow matching and the dedicated Alpha decoder under matched SFT conditions. Compared with LayerEditor w/o C-EPD, this variant lowers visual quality from 5.12 to 4.65 and increases RGBA L1 from 0.0482 to 0.0642 (Table 3, p. 9). These results support Direct’s contribution to visual coherence and layer reconstruction even without distillation. Removing Direct also eliminates the query state required to apply C-EPD through the current single-image teacher interface.

Table 3. Ablations on Crello-Testset (p. 9).

ModelEdit Cons. ↑Visual Qual. ↑Layer Seg. ↑RGBA L1 ↓
LayerEditor6.146.356.830.0350
w/o C-EPD4.675.125.010.0482
w/o Direct latent4.354.654.960.0642
Base Model3.934.133.770.0735
AuthorsOpenReview ↗

3. Learning Alpha prediction and boundary quality (Q3)

Thank you for examining the Alpha pathway carefully. LayerEditor explicitly includes RGB and Alpha tokens for each layer and trains both with flow-matching supervision, allowing the model to learn layer appearance and Alpha-mask prediction jointly in a supervised manner. In-Context RGB–Alpha Token Modeling coordinates appearance and opacity through joint attention, while layer-aware RoPE identifies each RGB–Alpha pair and preserves shared spatial coordinates (Section 3.1, pp. 3–4).

The Alpha decoder is trained through single-layer reconstruction, full-document supervision, and latent-noise adaptation (Appendix A.2, pp. 14–15; Table 7, p. 15). Reconstruction and perceptual losses supervise opacity and rendered appearance, while PatchGAN loss during full-document training encourages sharper boundaries. During C-EPD, both decoders remain frozen, but composition gradients pass through them to the student’s RGB and Alpha latents. Table 3 (p. 9) compares decoding approaches with Direct retained. Under SFT, adding the dedicated decoder to Base Model yields w/o C-EPD, improving visual quality from 4.13 to 5.12 and reducing RGBA L1 by 34.4%. Under full C-EPD, the dedicated decoder reduces RGBA L1 by 39.4%, from 0.0578 to 0.0350, compared with averaging the pretrained RGB decoder’s output channels. The qualitative examples also support these improvements, showing finer layer boundaries and fewer edge artifacts after composition.

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

LayerEditor adapts a pretrained image editor to jointly generate a Direct canvas and an ordered stack of RGB and alpha fields. The manuscript describes C-EPD as teacher velocity supervision for the Direct field together with pixel-space supervision of the composition decoded from local clean layer estimates. A separately trained alpha decoder supports opacity reconstruction. LayerEditor-100K supplies paired editing and layer supervision; experiments cover two editing benchmarks, conditional layer organization, and image-to-layer decomposition.

Critical strengths
  1. The task and reported system results are useful. Joint editing and conditional layer generation address a practical workflow, and the evaluation includes both native-design and flattened-edit sources. Reported Composite SSIM/FID and alpha agreement improve over the evaluated alternatives, and the held-out I2L experiment provides additional evidence beyond the main editing task.

  2. The manuscript describes the conditioning and evaluation boundaries carefully. Source-document splits keep derived edits together, reference pixels are withheld from generation, and count-only and layer-description-conditioned comparison groups are distinguished. The reported ablations support using the two teacher constraints together, although their interpretation depends on resolving the implementation issue below.

Critical weaknesses
  1. The supplied training artifact does not establish an implementation of the central C-EPD objective described in the manuscript. Equations 3–4 and Appendix A.4 decode local clean RGB and alpha estimates, compose in pixel space on white, and retain gradients through both decoders to their inputs. In the supplied archive, scripts/train.sh enables the direct_opd path with predicted-mask composite guidance. Its _composite_teacher_transition_loss instead combines RGB transition latents, obtains mask visibility through the standard RGB VAE under torch.no_grad(), and compares the latent composite with a teacher transition latent. The Direct loss is also computed between transition states, introducing a step-size-squared weighting relative to velocity MSE. This could be an earlier code version, but the connection between the stated method, supplied implementation, and reported tables is currently unresolved. Because compositional distillation is a central technical contribution, this affects the interpretation and reproducibility of its ablations.

  2. Practical downstream reuse is not yet established by the principal metrics. Reconstruction and alpha agreement in the original layout do not directly test moving, scaling, removing, or recombining the outputs. Figure 1 illustrates subsequent edits, and Appendix E acknowledges the lack of systematic manipulation tests. The single-VLM rubric has not been calibrated against human judgments, while unpremultiplied RGBA L1 can also change with invisible RGB storage, as Appendix C.2 notes. A small operation-based evaluation would connect the reported improvements more directly to the central reusable-layer motivation.

Overall assessment (updated after rebuttal). I now lean toward weak acceptance. The authors explained that the original archive contained an earlier latent-transition variant and provided the corrected implementation of the pixel-space C-EPD objective used for Tables 1–3. They also added a method-blind human evaluation with multi-round layer reuse, which directly addresses the downstream-reuse concern. My remaining reservation is the release of LayerEditor-100K, the benchmarks and the trained weights, which is described as pending approvals rather than committed; a firm release plan would further strengthen the contribution.

Questions for authors
  1. Which exact code version and configuration produced Tables 1–3? Please reconcile the supplied latent-transition composition, detached mask visibility, and transition-state loss with the manuscript's pixel-space local-clean composition and alpha gradients. If the archive is an older variant, could the implementation used for the reported results be provided?

  2. Can you provide a small, fixed-operation layer-reuse evaluation, or clearly delimit the current reuse claim to the demonstrated qualitative examples and original-layout reconstruction? A small human check of the main rubric conclusions would also help.

  3. What are the release or access plans for the trained LoRA/alpha decoder, LayerEditor-100K, and the benchmarks? If redistribution is constrained, what reproducibility route will be available with the implementation used in the paper?

4 comments
AuthorsOpenReview ↗

1. Corrected code implementation and correspondence with the reported results (W1, Q1)

Sorry for the confusion. The original archive mistakenly included an earlier latent-transition variant. We have submitted the corrected LayerEditor implementation. Tables 1–2 and the corresponding C-EPD ablations in Table 3 use the pixel-space objective in Eqs. (3)–(4) and Appendix A.4 (pp. 16–17), consistent with the updated code.

The differences concern both the supervision target and the gradient path. The earlier variant combined RGB latents after one denoising step. The student layer states and the teacher’s Direct state contain different noise realizations and have different starting values. Aligning the combined layer states with the teacher’s Direct state therefore includes residual-noise and starting-state discrepancies in the error. The earlier variant computed Mask visibility under no_grad, so the composition loss could not directly optimize Alpha predictions. In the corrected opd_clean_pixel.py, we first estimate and decode local clean RGB and Alpha latents, then composite the decoded layers and compare the result with the decoded teacher clean target. The composition loss operates in rendered image space, with gradients reaching both student RGB and Alpha predictions through the frozen decoders. The Direct term matches velocities without the squared step-size factor.

AuthorsOpenReview ↗

2. Human evaluation and practical usability of generated layers (W2, Q2)

We thank the reviewer for the constructive suggestion. We conducted a method-blind human evaluation, randomly sampling 500 cases from Crello-Testset and ScaleEdit-Testset and inviting 50 evaluators with professional design backgrounds to review all cases. Each case includes the source image, the layered editing result, and at least three rounds of layer reuse involving insertion, deletion, transformation, reordering, and text-layer editing. All methods underwent the same reuse tests. Evaluators rated edit consistency, visual quality, layer-content accuracy, layer-composition accuracy, and editing reusability on a 1–10 scale. A score of 10 indicates that the result fully meets the requirements, 7 indicates a usable result with only minor defects, 5 indicates partial success with noticeable problems, 3 indicates a largely unusable result, and 1 indicates complete failure. The evaluation protocol is described in Appendix C.4 (p. 23). As shown in Table 5 (p. 10), LayerEditor achieves the highest scores across all five dimensions. LayerEditor’s edit-consistency score is 7.17, compared with 5.97 for Seedream, the strongest baseline on this dimension. LayerEditor’s reusability score is 7.97, compared with 6.13 for LayerEditor w/o C-EPD and 4.81 for FR + QIL-Control, the strongest external baseline on reusability.

We found that competing methods frequently produce unintended mixing of content across layers and blurred layer boundaries, which hinder subsequent reuse and contribute to lower scores. We also added three figures showing how layer reuse supports subsequent controllable editing. Figure 8 in the main text (p. 10) demonstrates independent manipulation of individual layers, while Figures 15–16 in the supplementary material (pp. 27–28) show editing results for the same source image under different requested layer counts, followed by multiple rounds of layer reuse and editing. The reuse examples and human ratings show that LayerEditor can fulfill the initial editing request and support subsequent independent control of the generated layers. Human ratings assess the outputs after manipulation and complement per-layer RGBA L1. Appendix C.2 (p. 20) specifies how RGBA L1 treats stored RGB values.

Table 5. Human evaluation of layered editing (p. 10). Scores range from 1 to 10, with higher values indicating better results. Edit., VQ., Cont., Comp., and Reuse. denote edit consistency, visual quality, layer-content accuracy, layer-composition accuracy, and editing reusability, respectively.

MethodEdit. ↑VQ. ↑Cont. ↑Comp. ↑Reuse. ↑
Seedream5.975.575.605.174.80
FR + QIL5.636.072.605.573.50
FR + QIL-Decomp-FT3.973.274.204.233.88
FR + QIL-Control4.124.254.233.874.81
LayerEditor w/o C-EPD5.836.436.176.536.13
LayerEditor7.177.337.677.977.97
AuthorsOpenReview ↗

3. Resource release and reproducibility during the approval process (Q3)

We are actively preparing the open release, and the relevant approvals are in progress. We plan to release the code, model weights, including the LoRA adapter and dedicated Alpha decoder, and datasets, including LayerEditor-100K and the benchmarks, in stages as the respective approvals are granted.

To support reproducibility while these approvals are pending, the updated code package includes training, inference, and evaluation code, together with example source images, editing targets, ordered RGBA layers, structured prompts, and layer descriptions in sample_data/. Appendix B (pp. 17–18) details the data-construction and quality-control procedures. These examples and instructions also help readers run the implementation with compatible data they can obtain independently.

Dear Reviewer TZwD,

Thank you very much for your time and valuable feedback.

We understand that you may be busy. With the discussion deadline approaching and only one day remaining, we would appreciate it if you could review our responses and the revised manuscript.

Please let us know if any of your concerns remain unresolved. We would be very happy to address any further questions you may have.

Thank you for your attention.

Best regards, Authors of #7343

Comments on the paper

5 comments
AuthorsOpenReview ↗

Response to Common Concerns and Summary of Revisions

We thank all reviewers for their constructive feedback and thoughtful engagement with our work. We are grateful that the reviewers identified several consistent strengths in our submission, which affirm the value of our research. The comments have been invaluable in helping us improve our manuscript.

Strengths: LayerEditor introduces a unified model for image editing and content-controllable RGBA layer generation. Our proposed In-Context RGB–Alpha Token Modeling, dedicated Alpha decoder, and C-EPD jointly improve layer appearance and transparency while preserving editing capability. Extensive experiments demonstrate that LayerEditor maintains strong editing fidelity while generating accurate layers.

In response to common concerns, we strengthened the manuscript in two main areas:

1. Multi-round layer reuse evaluation. We conducted a method-blind user study with 50 professional designers evaluating 500 cases. LayerEditor leads all five dimensions in Table 5 (p. 10), supporting editing quality and subsequent reusability. We also added qualitative demonstrations in Figure 8 (p. 10) and Figures 15–16 (pp. 27–28) to demonstrate independent layer manipulation and multi-round reuse without image regeneration.

2. Additional ablation studies. We added a w/o Direct latent variant and two off-policy query-selection ablative variants in Tables 3–4 (p. 9). The results support Direct’s contribution to visual quality and layer reconstruction, and show that current-student on-policy queries better preserve editing fidelity and layer quality. We also clarified how the requested layer count and ordered layer-content descriptions work together to control decomposition granularity and semantic grouping. Table 2 (p. 8) evaluates layer organization under a fixed editing intent, while Appendix D.1, Table 8 (p. 24) evaluates whether revised instructions produce the corresponding edits and layer structures.

The additions and clarifications appear in:

  • Section 4.5, Tables 3–4 (p. 9): Direct-latent ablation and on-policy/off-policy comparisons.

  • Section 4.6, Table 5 and Figure 8 (p. 10): Human evaluation and layer-reuse demonstration.

  • Appendix C.2 (p. 21): Implementation details for the additional controlled experiments.

  • Appendix C.4 (p. 23): Human evaluation and reuse protocol.

  • Figures 15–16 (pp. 27–28): Alternative layer organizations and multi-round manipulation.

AuthorsOpenReview ↗

1. Key Designs and On-Policy Guidance for Preserving Editing Capability (W1, Q1)

LayerEditor unifies source-image editing with text-controlled generation of reusable RGBA layers, allowing users to specify both layer content and count. QIL models variable-sized RGBA stacks, but its demonstrated content-editing workflow invokes a separate Qwen-Image-Edit model after decomposition. MRT supports layered generation and editing, but its decomposition requires spatial regions supplied by a detector or manual annotation. LayerEditor jointly performs image editing and semantic layer organization through instructions. For example, users can change a poster’s wording while specifying whether people or decorations share a layer or remain independently editable (Figures 8 and 15–16, pp. 10 and 27–28). The prompt controls both the visual edit and the grouping needed for subsequent reuse.

The central challenge is to preserve a pretrained editor’s editing capability while expanding a single-image output into coordinated RGB–Alpha layers. Individual layers must follow the requested semantic grouping, handle transparency and occlusion, and remain faithful to the edit after composition. A single-image teacher cannot directly supervise these layer-specific outputs. Our contribution is to extend image-level on-policy guidance to joint canvas-and-layer generation. In-Context RGB–Alpha Token Modeling coordinates appearance and opacity, while the dedicated Alpha decoder reconstructs transparency. C-EPD samples queries from the current student’s rollout, allowing supervision to follow the evolving denoising states and errors during layered adaptation. At selected detached queries, we recompute student predictions with gradients. The frozen teacher receives the Direct latent and provides a velocity target for Direct and a decoded clean-image target for the RGBA composition. Composition gradients pass through both frozen decoders to student RGB and Alpha latents (Appendices A.3–A.4, pp. 15–17). This design makes the single-image editing prior applicable to the visible result of an entire layer stack, without requiring a layer-wise teacher.

We evaluate query selection using SFT Rollout OPD, which samples queries from a frozen SFT model, and GT Noised OPD, which adds noise to ground-truth joint latents. Both controls share C-EPD’s initialization, data, teacher, decoder, losses and weights, query noise levels and count, and training budget. Only query construction changes (Appendix C.2, p. 21). C-EPD outperforms both controls across all four metrics (Table 4, p. 9). Relative to SFT Rollout OPD, edit consistency increases by 9.3%, from 5.62 to 6.14, visual quality improves by 16.1%, and RGBA L1 decreases by 21.0%, from 0.0443 to 0.0350. These results support current-student queries within C-EPD’s Direct–Composite guidance for preserving editing fidelity and layer quality. Figure 7 (p. 9) illustrates the editing improvements, while Figures 8 and 15–16 demonstrate independent manipulation and multi-round reuse.

Table 4. On- and off-policy OPD comparisons on Crello-Testset (p. 9). Edit Cons., Visual Qual., and Layer Seg. are 1–10 rubric scores for edit consistency, visual quality, and layer segmentation. Lower RGBA L1 indicates better layer reconstruction.

ModelEdit Cons. ↑Visual Qual. ↑Layer Seg. ↑RGBA L1 ↓
w/o Comp. OPD (Vanilla on-policy OPD)4.984.334.970.0492
SFT Rollout off-policy OPD5.625.476.610.0443
GT Noised off-policy OPD5.445.336.320.0471
LayerEditor6.146.356.830.0350
AuthorsOpenReview ↗

2. Human evaluation and multi-round layer reuse (W2, Q2)

We thank the reviewer for the constructive suggestion. We conducted a method-blind study on 500 cases randomly sampled from Crello-Testset and ScaleEdit-Testset. All 50 evaluators with professional design backgrounds reviewed all cases. Each case includes the source image, layered editing output, and at least three successive rounds of reuse involving insertion, deletion, transformation, reordering, and text-layer editing. All methods used the same operation sequences and parameters, with previous edits retained at each round. We recomposed the modified stacks without regeneration or repair. Evaluators rated edit consistency, visual quality, layer-content accuracy, layer-composition accuracy, and editing reusability on a 1–10 scale (Section 4.6, p. 10; Appendix C.4, p. 23).

Table 5 (p. 10) shows that LayerEditor leads all five dimensions. LayerEditor’s reuse score is 7.97, compared with 6.13 for LayerEditor w/o C-EPD and 4.81 for the strongest external baseline, FR + QIL-Control. Figure 8 (p. 10) shows independent manipulation of the turkey and basket, exposing content previously hidden in the original composition. Figures 15–16 (pp. 27–28) show alternative requested layer organizations followed by successive rounds of reuse for each generated stack. The human ratings directly assess the outputs after layer manipulation.

Table 5. Human evaluation of layered editing (p. 10). Scores range from 1 to 10, with higher values indicating better results. Edit., VQ., Cont., Comp., and Reuse. denote edit consistency, visual quality, layer-content accuracy, layer-composition accuracy, and editing reusability, respectively.

MethodEdit. ↑VQ. ↑Cont. ↑Comp. ↑Reuse. ↑
Seedream5.975.575.605.174.80
FR + QIL5.636.072.605.573.50
FR + QIL-Decomp-FT3.973.274.204.233.88
FR + QIL-Control4.124.254.233.874.81
LayerEditor w/o C-EPD5.836.436.176.536.13
LayerEditor7.177.337.677.977.97

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.