acceptodds
Under review as a conference paper at ICLR 2027

A Unified and Controllable Framework for Layered Image Generation with Visual Effects

Abstract

Recent image generation models produce impressive composites, but often fail to preserve the identity of user-provided content when editing specific elements: the surrounding scene may shift, and even the edited object’s appearance can drift from the original. Layered representations offer a natural remedy—they allow users to independently manipulate individual elements—but existing layered methods typically produce transparent foregrounds without realistic visual effects such as shadows and reflections, forcing the use of a second harmonization model after every edit, which in turn reintroduces drift. To overcome these limitations, we present LASAGNA, which generates a photorealistic background and an RGBA foreground with compelling visual effects in a single forward pass. By treating object-associated visual effects as part of the foreground layer, LASAGNA supports the dominant class of consumer edits (e.g., translation, scaling, recoloring, duplication) via alpha compositing alone, without invoking any model post-edit, thereby eliminating the identity drift inherent to cascade editing pipelines. This single-pass design contrasts with prior layered methods that rely on separate expert models for each task; LASAGNA handles diverse conditioning inputs—text prompts, foreground, background, and location masks—within a unified architecture. We further release two community resources: LASAGNA-48K, the first public dataset of 48K layered image triplets with physically grounded visual effects (curated via a VLM-based filter trained on 30K human-labeled examples), and LASAGNABench, the first standardized benchmark for layer-centric generation and editing, comprising 242 expert-annotated samples across six diverse sources. Experiments show that LASAGNA outperforms both general-purpose editors (FLUX, Qwen-Image-Edit, gpt-image-1) and prior layered methods (LayerDiffuse) across three generation modes, and supports a wide range of post-edits without any neural re-inference.

Primary area:
generative models
Keywords:
Layered Image Generation, Controllable Image Editing, Visual Effects, Alpha Compositing, Multimodal Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_J5pBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents LASAGNA, a unified framework for layered image generation with visual effects. The authors argue that existing layered methods produce transparent foregrounds without realistic shadows/reflections, forcing a harmonization model after every edit and thereby reintroducing identity drift. LASAGNA generates a photorealistic background and an RGBA foreground with object-associated visual effects in a single forward pass, treating effects as part of the foreground layer so that consumer edits (translation, scaling, recoloring, duplication) can be done by alpha compositing alone without model re-inference. A single DiT-based architecture with type/IO/position embeddings handles three modes—background-conditioned foreground generation, foreground-conditioned background generation, and text-to-all layer generation—under diverse conditioning inputs (text, FG, BG, masks). The authors also release LASAGNA-48K, a 48K layered triplet dataset filtered by a VLM-based curator, and LASAGNABENCH, a 242-sample expert-annotated benchmark across six sources. Experiments compare against FLUX.1, Qwen-Image-Edit, gpt-image-1, and LayerDiffuse, plus ablations, user study, and additional benchmarks.

Critical strengths
  1. The core idea—binding scene-consistent visual effects into the foreground alpha layer—is simple, well-motivated, and practically valuable, enabling zero-inference post-editing and eliminating cascade drift.
  2. The unified architecture handles three generation modes and heterogeneous conditioning within one model, contrasting favorably with prior per-task expert models; the type/IO embedding interface is a clean design.
  3. Community contributions are substantial: the first public layered dataset with visual effects and the first standardized layer-centric benchmark, both with documented construction pipelines and human verification.
  4. Experiments are broad: comparisons against both general models and the closest layered baseline, three modes, GPT-based identity/instruction metrics, a user study, and a paradigm-level necessity analysis (instruct vs. cascaded layer vs. layer+effects).
  5. The 2B model outperforms much larger general models on several metrics, and the runtime/editing advantage is clearly demonstrated.
Critical weaknesses
  1. The dataset is built by filtering LayerDecomp outputs with a learned curator; residual decomposition artifacts and the curator's 72.3% recall may propagate into training and benchmark labels, and the benchmark's "ground-truth" BG/FG for public sources is itself model-generated rather than captured.
  2. Evaluation relies heavily on FID/CLIP-FID and GPT-4o-based scores; the GPT judge may be biased toward the authors' outputs, and no human evaluation of visual-effect physical plausibility is reported beyond a preference study.
  3. The method is evaluated mainly in the "FG over BG" single-foreground setting; multi-object interactions, occlusion, and complex scene geometry are acknowledged as future work, limiting claims of generality.
  4. Comparisons with general models use their task-specific expert variants, which is reasonable but means LASAGNA's advantage in "unified" capability is not isolated from raw quality; LayerDiffuse comparison is limited to CLIP-FID.
  5. Training details (20K iterations, 2B DiT, RGBA-VAE fine-tuning) and data scale are modest; ablation of training data, prompt formulation, and joint training is relegated to the appendix, and sensitivity to curator/filter thresholds is unclear.
Questions for authors
  1. How much do LayerDecomp artifacts and curator errors propagate into LASAGNA-48K and the benchmark labels? Can you report metrics on human-captured in-house pairs separately?
  2. How reliable is the GPT-4o judge, and can you provide human evaluation of shadow/reflection physical plausibility and identity preservation?
  3. How does LASAGNA behave with multiple foreground objects, occlusions, and interacting shadows? What breaks first?
  4. Since baselines use expert variants, can you report a unified-model comparison or a compute-matched comparison to isolate the benefit of the unified interface?
  5. How sensitive are results to the RGBA-VAE fine-tuning and the 20K-iteration training budget? Would a larger data scale change the conclusions?
2 comments

We thank the reviewer for recognizing that our core idea (binding scene-consistent visual effects into the foreground alpha layer) is simple, well-motivated, and practically valuable, that the Type/IO interface is a clean design unifying three modes in one model, that our dataset and benchmark are substantial community contributions, that our experiments are broad, and that our 2B model outperforms much larger general models.

Q1: How much do LayerDecomp artifacts and curator errors (72.3% recall) propagate into LASAGNA-48K and the benchmark labels?

R: Very little. For the curator, the 88.8% precision matters more than the 72.3% recall: recall is the fraction of good cases retained, whereas precision is the probability that a sample admitted into the dataset is good. We deliberately trade a little recall for higher precision, so curator errors mainly discard good samples rather than admit bad ones.

Moreover, the benchmark labels of the public sources are not all model-generated: all FGs are human-extracted and the BGs of the two camera sets are real captures; only the BGs of the public sources are model-generated, and all of them passed human verification. Overall, 66.5% of the benchmark's FG/BG layers are produced entirely by humans without any model involvement (all 242 FGs plus the 80 camera-captured BGs, out of 484 layers; Table 10). This high-quality benchmark ensures that the model is iterated in the right direction even if the training set contains slight noise. For the training data, strict filtering (a curator trained on 30K human annotations with hard negatives, plus a 32B VLM) leaves >90% of samples judged good; see Q1 of Reviewer fZAY. Results on the human-captured in-house pairs are reported in Q2 below.

Q2: Reliability of the GPT-4o judge; human evaluation of shadow/reflection physical plausibility and identity preservation; metrics on human-captured in-house pairs.

R: Human evaluation confirms our GPT-4o-based results, both overall and on the human-captured in-house pairs.

  1. Human evaluation. We conducted a comprehensive blinded user study with 1,878 criterion-level votes (Appendix D, Table 6) along three dimensions:

    • Instruction Following (IF): compliance with text and task-specific spatial constraints;
    • Identity Preservation (IP): preservation of the original BG in FG_Gen or of the FG object in BG_Gen;
    • Harmonization and Quality (HQ): natural integration of boundaries, lighting, perspective, and shadows, together with perceptual quality and artifacts.

    HQ evaluates shadow/reflection physical plausibility, and IP evaluates identity preservation. Our model is the most favored, exceeding the best baseline in Task Avg by 35.6 points in FG_Gen and 38.2 in BG_Gen (see also Q3 of Reviewer nwuN). This ranking is consistent with the GPT-4o scores, showing that the GPT-4o judge is reliable and not biased toward our outputs.

  2. Metrics on human-captured in-house pairs. From the user study, we separately extract the statistics of the in-house samples: 40 camera-captured samples across the two tasks, with 549 valid votes.

Note: BWS stands for Best-Worst Scaling, where scores range from -100 to +100, and higher scores indicate stronger human preference (Appendix D, Table 6) .

FG_Gen (259 votes; IF/IP/HQ: 94/84/81)

ModelIF BWSIP BWSHQ BWSTask Avg
Ours+69.3+41.3+28.8+46.5
Qwen-17.9+4.4+0.4-4.4
FLUX-19.0+40.1-44.4-7.8
GPT-32.4-85.7+15.2-34.3

BG_Gen (290 votes; IF/IP/HQ: 98/97/95)

ModelIF BWSIP BWSHQ BWSTask Avg
Ours+36.2+44.5+25.6+35.4
GPT+38.1-85.4+45.6-0.6
Qwen-27.7+21.5+3.5-0.9
FLUX-46.6+19.4-74.7-34.0

Our model achieves the highest Task Avg in both tasks by a wide margin (+46.5 vs. -4.4 and +35.4 vs. -0.6 for the best baseline); the conclusions on real-captured data remain consistent with our original findings.

Q3: How does LASAGNA behave with multiple foreground objects, occlusions, and interacting shadows? What breaks first?

R: LASAGNA supports multiple foreground objects, occlusions, and interacting shadows; see Fig. S1 and Fig. S2 in the newly updated supplementary material. Occlusions and interacting shadows are harder than multiple foreground objects, yet the results remain strong although such cases were never deliberately collected for training. Please also see Q1 of Reviewer XXLt (multi-object layers in a multi-turn manner, Figure 8) and Q2 of Reviewer fZAY (fine hair, fur, and semi-transparency are handled well).

Q4: Since baselines use expert variants, can you report a unified-model comparison or a compute-matched comparison to isolate the benefit of the unified interface?

R: Yes, this is already isolated in Table 9 (third row, "Separate task"): under the same training samples, the unified model outperforms separately trained single-task models in all three modes. The gain therefore comes from the unified interface itself.

Q5: More evaluation metrics for the LayerDiffuse comparison (currently limited to CLIP-FID).

R: Our model also significantly outperforms LayerDiffuse on GPT scores. Following the protocol in our paper, we additionally report them on FG_Gen and BG_Gen:

ModelFG_Gen IF ↑FG_Gen IP ↑FG_Gen AvgBG_Gen IF ↑BG_Gen IP ↑BG_Gen Avg
LayerDiffuse7.96.27.15.85.35.6
Ours9.279.299.39.348.679.0

This is consistent with Figure 6 and Table 4. We additionally compare with BFS (SIGGRAPH 2026), a recent dedicated FG_Gen model, where our FG layers are also clearly better (FG FID 151.7 vs. 234.9); see Q2 of Reviewer nwuN and Fig. S3 in the newly updated supplementary material.

Q6: How sensitive are results to the RGBA-VAE fine-tuning and the 20K-iteration training budget? Would a larger data scale change the conclusions?

R: Our conclusions are stable: the reported results have converged within the 20K-iteration budget, and a larger data scale would only strengthen them. The RGBA-VAE is important, but it is a general component well validated by the community and outside the scope of our paper, so we did not ablate it specifically. Larger and more diverse data should make the model stronger and may further benefit the challenging cases discussed above.

We hope this addresses your concerns; please feel free to ask any further questions.

Reviewer_nwuNOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Lasagna, a novel framework designed to generate images as a composition of foreground and background layers. The proposed framework supports 3 generation modes: FG_Gen, BG_Gen, and Text2All, each of which supports background-conditioned foreground layer generation, foreground-conditioned background generation, and text-to-all layer generation. To achieve this, the proposed method adapts a pretrained DiT model. The adaptation is straightforward as the DiT model is readily extendible to take various types of input conditions by simply concatenating them as tokens. More importantly, the paper proposes novel datasets named Lasagna-48K, and LasagnaBench for training and evaluation of layer generation. The Lasagna-48K dataset is synthetically generated using a data construction pipeline. The pipeline consists of 4 stages: non-occluded mask selection, LayerDecomp decomposition, data filtering, and captioning. The LasagnaBench dataset provides both synthetically generated samples and real-world samples.

Critical strengths

The paper is clearly written and easy to read.

LASAGNA-48K and LASAGNABENCH could be valuable resources for future research.

The comparison shows that the proposed method outperforms LayerDiffuse.

The proposed framework is flexible, supporting three generation modes.

Critical weaknesses

Unfortunately, the technical contributions are rather limited. Adapting a pretrained DiT to layer generation is straightforward as the DiT model already supports flexible inputs. The dataset synthesis pipeline is also a rather simple combination of existing methods and manual curation.

Although the paper emphasizes generating foreground layers with visual effects, it does not introduce a dedicated mechanism for modeling these effects or enforcing their consistency with the background. This capability appears to arise primarily from training on effect-aware layered data, rather than from a specific architectural design or training objective, which further limits the methodological novelty.

The evaluation is also limited. The paper evaluates several general-purpose models, but LayerDiffuse is the only dedicated layered generation baseline. There have recently been introduced several layer generation and decomposition approaches to support layered images, which must be included in the comparisons.

  • Kang et al., BFS: Back-to-Front Layered Image Synthesis via Knowledge Transfer, SIGGRAPH 2026
  • Qwen-Image-Layered, https://qwen.ai/blog?id=qwen-image-layered
  • Suzuki et al., LayerD: Decomposing Raster Graphic Designs into Layers, ICCV 2025
  • Kang et al., LayeringDiff: Layered Image Synthesis via Generation, then Disassembly with Generative Knowledge, arXiv 2025 A particularly relevant method among the papers listed above is BFS, which also generates a foreground layer containing shadows and reflections conditioned on a background. If implementations or checkpoints are unavailable, a careful discussion and at least a qualitative comparison using published examples would help establish the contribution, while acknowledging differences in inputs and evaluation settings.

The paper includes a human evaluation of harmonization and quality, as well as qualitative challenging cases, so these aspects are not entirely unexamined. However, FID, CLIP-FID, and aggregate preference scores provide limited evidence about specific relationships such as object scale, contact with supporting surfaces, illumination, and shadow/reflection geometry. More targeted evaluation would strengthen these claims.

The paper reports source distributions and examples, but gives less detail about coverage across object categories, scene complexity, transparency, and types of visual effects. Such statistics, together with results broken down by these factors, would make generalization easier to assess.

Questions for authors

Please address the concerns discussed above.

5 comments

We thank the reviewer for recognizing that the paper is clearly written, that LASAGNA-48K and LASAGNABENCH are valuable resources, that our method outperforms LayerDiffuse, and that our framework flexibly supports three generation modes.

Q1: Limited technical contributions (straightforward DiT adaptation; no dedicated mechanism for modeling visual effects; dataset pipeline as a simple combination of existing methods and manual curation).

R: Our core contribution is formulating layer generation as a unified, configurable set-to-set generation problem, where foregrounds with visual effects, backgrounds, and composites can each serve as conditioning inputs or generation targets within a shared model. We agree that adapting a pretrained DiT to heterogeneous inputs is not, by itself, the central contribution.

  1. Unified formulation. Our Type and IO embeddings encode complementary information: each frame's semantic role and whether it is provided or generated. The same model thus supports FG_Gen, BG_Gen, and Text2All through different input/output configurations, without independently trained expert models (e.g., LayerDiffuse needs three separate models) or task-specific models (e.g., BFS is designed specifically for FG_Gen). As shown in Figure 9, we can extend to a new task while reusing the model's existing capability, simply by changing frame-role assignments. Moreover, the unified model improves composite CLIP-FID over separately trained task-specific models in all three modes (11.0→10.3, 15.6→14.6, 19.2→17.8; Table 9).
  2. Visual effects without a dedicated mechanism. This is by design: binding visual effects to the foreground layer and jointly denoising the composite, BG, and FG in one shared attention sequence lets the model learn effects and their consistency with the background directly, with no extra module or loss; post-edits then reduce to pure alpha compositing without re-inference. Empirically, this simple data-driven design is effective: our FG layers carry scene-consistent shadows and reflections across diverse real scenes and are clearly better than those of dedicated layered baselines (Table 4 and Q2; Fig. S3 in the newly updated supplementary material), suggesting that explicitly supervising visual effects with real-world layered data is the decisive factor.
  3. Dataset pipeline. The key component is not the combination of existing tools but our curator, trained on 30K samples that we manually annotated under a high standard (88.8% precision). Its practical value is substantial: curating layers with faithful visual effects is a very fine-grained requirement that no combination of existing methods or off-the-shelf tools can meet without it. The result is the first public layered dataset with visual effects.

We will revise the paper to better distinguish the novelty of the unified formulation from the individual architectural components.

Q2: More evaluations (BFS, Qwen-Image-Layered, LayerD, LayeringDiff).

R: We add a quantitative comparison with BFS, the most relevant method, on LASAGNABENCH: our FG layers are clearly better (FG FID 151.7 vs. 234.9), while the composites are comparable. The other three methods cannot be compared: LayeringDiff has not released checkpoints; LayerD is a decomposition rather than a generation model; and Qwen-Image-Layered is also mainly designed for decomposition (we tried its actual Text-to-RGBA code, which does not support generating a new layer). Comparison with BFS (a dedicated FG_Gen model):

ModelFG FID↓FG CLIP-FID↓Composite FID↓Composite CLIP-FID↓Re-composite FID↓Re-composite CLIP-FID↓
BFS234.940.772.19.374.09.9
Ours151.728.272.09.771.69.7

On composite and re-composite images (the generated RGBA FG alpha-composited onto the input BG), the two are comparable, while our FGs are clearly better. As visualized in Fig. S3 in the newly updated supplementary material, BFS generates high-quality objects, but its FG layers sometimes lack the corresponding visual effects, whereas ours consistently carry shadows and reflections that match the scene. This explains why the composite metrics are close while our FG layers are much better, and it is exactly what makes our layers directly editable by alpha compositing.

We also compare with ART [1]: on the multi-layer cases from the ART paper, our model achieves comparable layer quality with more realistic shadows, without any specific optimization (Fig. S1 in the newly updated supplementary material; see Q3 of Reviewer XXLt).

Q3: More targeted evaluation of object scale, contact with supporting surfaces, illumination, and shadow/reflection geometry.

R: Our user study already assesses these aspects at the criterion level, rather than by aggregate preference. It is a comprehensive blinded study with 1,878 criterion-level votes (Appendix D, Table 6) along three dimensions:

  • Instruction Following (IF): compliance with text and task-specific spatial constraints, covering object scale and placement;
  • Identity Preservation (IP): preservation of the original BG in FG_Gen or of the FG object in BG_Gen;
  • Harmonization and Quality (HQ): natural integration of boundaries, lighting, perspective, and shadows, together with perceptual quality and artifacts, covering contact with supporting surfaces, illumination, and shadow/reflection geometry.

Targeted hard cases are further examined in Appendix G (Figure 15), and the camera-captured in-house subset in Q2 of Reviewer J5pB. The conclusion of the user study is consistent with FID and CLIP-FID: our model achieves the best results among existing public models.

Q4: Data coverage across object categories, scene complexity, transparency, and types of visual effects.

R: Our benchmark has broad coverage along all of these factors.

  1. LASAGNABENCH. Using caption-derived semantic statistics on the 242 samples, we identify 10 broad object groups spanning animals, people, vehicles, household objects, and plants across diverse scenes. By the number of distinct background element families mentioned, 55.0% / 36.4% / 8.7% of samples exhibit low / medium / high scene complexity. Transparent/clear materials or glass components are mentioned in 14 samples (5.8%), and 91 samples (37.6%) explicitly mention visual effects, including reflections, shadows, blur, lighting, gloss, and smoke/haze.
  2. LASAGNA-48K. Given its large size, we will add its detailed statistics, together with results broken down by these factors, in the revision.

[1] ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation, CVPR 2025

We hope this addresses your concerns; please feel free to ask any further questions.

Dear Reviewer nwuN,

Thank you again for your time and the constructive feedback on our work.

As the discussion phase is drawing to a close and the deadline approaches, we wanted to gently follow up to see if our previous responses and the newly added experiments have addressed your initial concerns.

If you have any further questions or if there are any remaining points you would like us to clarify, please do not hesitate to let us know. We are more than happy to provide any additional details before the deadline.

Thank you very much for your time and effort in reviewing our paper!

Best regards, Authors

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

The paper presents LASAGNA, a unified diffusion transformer (DiT) framework for layered image generation and editing that outputs a photorealistic background and an RGBA foreground explicitly embedding visual effects like shadows and reflections. By handling multiple modes (foreground generation, background generation, and text-to-all generation) within a single architecture via learnable Type and IO embeddings, LASAGNA facilitates consumer edit operations (such as translation, scaling, or duplication) using alpha compositing alone without re-inference. To train the model, the authors introduce LASAGNA-48K, a public dataset of 48K layered image triplets curated through a VLM filter trained on human-annotated examples. They also release LASAGNABENCH, a standardized benchmark of 242 expert-annotated samples.

Critical strengths
  1. Explicit Modeling of Visual Effects in Layers: Unlike prior transparent layer generation methods (e.g., LayerDiffuse) that output isolated sprites without scene-consistent shadows or reflections, LASAGNA binds these contextual effects to the foreground layer, enabling post-editing via simple alpha compositing without secondary harmonization models.

  2. Unified Architecture via Task-Aware Embeddings: The introduction of learnable Type and IO embeddings allows a single DiT backbone to fluidly switch between background-conditioned, foreground-conditioned, and text-to-all generation modes without changing model structure or weights.

  3. Valuable Public Resources: The release of both a large-scale training corpus (LASAGNA-48K) and a rigorous standardized evaluation benchmark (LASAGNABENCH spanning six diverse sources) addresses a notable gap in data availability for layer-centric generative research.

Critical weaknesses
  1. Ambiguities in Layer Decomp Dependency and Error Propagation: The data construction pipeline relies heavily on an automated decomposition model (LayerDecomp) to source training pairs, followed by a VLM filter. It remains under-explored how errors or artifacts introduced during automated decomposition, especially regarding complex overlapping shadows or semi-transparent reflections, affect the lower bounds of model fidelity.

  2. Limited Discussion on Boundary and Transparency Failures: While the paper showcases strong qualitative samples, edge matting and alpha channel precision for complex geometry (e.g., fine hair, fur, or semi-transparency intersecting with complex backgrounds) are critical failure points for layer decomposition models that are not thoroughly analyzed.

Questions for authors
  1. Decomposition Artifacts: How robust is the data curation pipeline to systematic errors from LayerDecomp in cases where shadows overlap with textured surfaces?

  2. Alpha Channel Quality: What is the quantitative precision of alpha matte boundary estimation in LASAGNA compared to ground-truth human annotations on LASAGNABENCH, particularly for thin structures?

3 comments

We thank the reviewer for recognizing our explicit modeling of visual effects in layers (enabling post-editing via simple alpha compositing without secondary harmonization models), our unified architecture that switches modes via Type and IO embeddings without changing model structure or weights, and the value of LASAGNA-48K and LASAGNABENCH as public resources.

Q1: Ambiguities in LayerDecomp dependency and error propagation; robustness of the data curation pipeline to systematic LayerDecomp errors where shadows overlap with textured surfaces.

R: Our curation pipeline is robust to LayerDecomp errors: its failure cases have no noticeable impact on the final dataset or our conclusions.

  1. Strict filtering of training data. We agree that LayerDecomp produces some artifacts, which is exactly why we invested heavily in filtering to minimize error propagation. Our curator is trained on 30K manually annotated samples under a high standard (Appendix K, Figure 19) with many hard negatives: e.g., the "tub" case (Figure 19, second row, first image) has an almost clean BG but is still labeled bad because slight visual effects of the tub remain. The curator achieves 88.8% precision and 72.3% recall, i.e., we may miss some good samples but are unlikely to keep bad ones. A 32B VLM then filters the remaining artifacts. As a result, >90% of LASAGNA-48K samples are judged good (line 352). In Figure 16, the dog (row 1), horse (row 2), and zebra (row 3) cases all lie on textured surfaces yet have well-decomposed visual effects.
  2. Human-annotated benchmark. All FGs are human-extracted and the BGs of the two camera sets are real captures; overall, 66.5% of the benchmark's FG/BG layers are produced entirely by humans without any model involvement (all 242 FGs plus the 80 camera-captured BGs, out of 484 layers; Table 10). This high-quality benchmark ensures that the model is iterated in the right direction even if the training set contains slight noise. Indeed, our conclusions hold on the camera-captured in-house pairs alone (see Q2 of Reviewer J5pB).

Q2: Limited discussion on boundary and transparency failures (fine hair, fur, semi-transparency intersecting with complex backgrounds).

R: Our model handles these boundary and transparency cases well. Fig. S2 in the newly updated supplementary material adds challenging cases covering occlusion ordering, effect–effect interaction, semi-transparency intersecting with complex backgrounds, fine hair, and fur. Notably, it achieves good results on fine hair and fur although it is not an expert matting model trained on deliberately collected matting data, and we did not specifically add semi-transparency data either; our data are real photos with prominent visual effects. We will add this discussion to the revision.

Q3: Alpha channel quality: quantitative precision of alpha matte boundary estimation against ground-truth human annotations on LASAGNABENCH, particularly for thin structures.

R: Alpha mattes in our setting cannot be directly compared against ground-truth human annotations: in a generation task, the generated object is not the same object as the GT, so its boundary differs and no ground truth exists for computing matting metrics. Matting-style precision is therefore not applicable to layer generation.

We hope this addresses your concerns; please feel free to ask any further questions.

Thank you for the detailed response and the additional experiments. Most of my concerns have been addressed, so I'm inclined to maintain my score.

Thank you for your time and for reviewing our response. We are glad that the additional experiments resolved your concerns. We sincerely appreciate your constructive feedback and your support.

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

This paper presents a unified diffusion transformer framework for layered image generation that outputs a photorealistic background and an RGBA foreground carrying object-bound visual effects (shadows, reflections) in a single forward pass. A set-to-set manner, in which frame roles are specified by Type and IO embeddings over a shared token sequence, unifies three modes—background-conditioned foreground generation, foreground-conditioned background generation, and text-to-all generation—within one 2B model. Because visual effects travel with the foreground layer, translation, scaling, recoloring, and duplication reduce to alpha compositing with no further neural inference, avoiding the identity drift of cascade pipelines. The authors additionally release LASAGNA-48K, a 48K-triplet dataset curated by a VLM-based filter, and LASAGNABench, a 242-sample expert-annotated benchmark. Across the three modes the method improves CLIP-FID, FID, and GPT-4o-based scores over FLUX, Qwen-Image-Edit, gpt-image-1, and LayerDiffuse, and wins a blinded human study by a wide margin.

Critical strengths
  1. This paper focus on a fundantmental research problem: how to build layer-wise generation models. The visual effects in each image layer is even a more important question, thus affecting the finally composed image.
  2. The authors bind object-associated visual effects to the RGBA foreground makes translation, scaling, recoloring and duplication pure alpha compositing, so post-edits need no neural re-inference and inherit no identity drift.
  3. The result is impressive, supressing some 12B image models.
Critical weaknesses

1, The framework is essentially one foreground over one background; multi-object layers, occlusion ordering, and effect–effect interactions are never evaluated 2. Visual-effect correctness is judged perceptually only; no metric checks geometry or physics (shadow direction, reflection location), so the claimed physically grounded claim is under-measured. 3. No comparsion with recent layer-wise model ART[1] and MRT[2].

[1]ART: Anonymous region transformer for variable multi-layer transparent image generation, CVPR2025 [2]Masked Region Transformer for Layered Image Generation and Editing at Scale, CVPR2026

Questions for authors

After an edit moves a foreground, its baked-in shadow and reflection no longer match the new location or lighting. How often does this visibly break, and is the intended remedy to re-run BG_Gen?

3 comments

We thank the reviewer for recognizing that our work tackles a fundamental problem (layer-wise generation with per-layer visual effects), that binding effects to the RGBA foreground makes post-edits pure alpha compositing with no re-inference or identity drift, and that our 2B model surpasses 12B image models.

Q1: The framework is essentially one foreground over one background; multi-object layers, occlusion ordering, and effect–effect interactions are never evaluated.

R: Our framework is not limited to one foreground over one background: it supports multi-object layers, occlusion ordering, and effect–effect interactions.

  1. Multi-object layers. We support them in a multi-turn manner (Figure 8); Fig. S1 in the newly updated supplementary material further shows four FG layers on one BG, each with its own shadow. A solid single-layer capability is the prerequisite for multiple layers. Single-shot multi-object generation is worth exploring in future work, but its data collection is clearly costlier than for a single object.
  2. More challenging scenarios. Fig. S1 and Fig. S2 in the newly updated supplementary material cover occlusion ordering and effect–effect interactions. The results are strong even though such cases were never deliberately collected for training (our data are real photos with prominent visual effects).

Our framework thus generalizes well beyond the single-foreground setting.

Q2: Visual-effect correctness is judged perceptually only; clarification of the "physically grounded" claim.

R: "Physically grounded" means that the visual effects are learned from real-world data, rather than from cartoon or synthetic data (as in most prior work) or computed by a renderer / 3D algorithm. Their correctness is also assessed by humans: the Harmonization and Quality (HQ) criterion of our blinded user study rates lighting, perspective, and shadows, and our model is the most favored overall (Appendix D, Table 6; see also Q2 of Reviewer J5pB). We will make this definition explicit in the revision.

Q3: Comparison with ART [1] and MRT [2].

R: On the multi-layer cases from the ART paper, our model achieves comparable layer quality without any specific optimization, and its layers additionally carry visual effects (Fig. S1 in the newly updated supplementary material): e.g., ART's bed has no shadow and appears to float in the composite, whereas each of our layers is grounded by its own shadow. Neither ART nor MRT currently provides publicly available checkpoints; we therefore re-run our model on the images used in the ART paper.

Moreover, ART and MRT focus on poster / graphic design, where images typically contain many tiny logo and text layers with almost no visual effects, whereas our model and datasets focus on realistic images with visual effects, a setting these methods do not cover.

We additionally compare with BFS (SIGGRAPH 2026), a recent layered generation method, on LASAGNABENCH, where our FG layers are clearly better (FG FID 151.7 vs. 234.9); see Q2 of Reviewer nwuN and Fig. S3 in the newly updated supplementary material.

Q4: After an edit moves a foreground, its baked-in shadow and reflection no longer match the new location or lighting. How often does this visibly break, and is the intended remedy to re-run BG_Gen?

R: Small local moves are the most common editing operation and what our design targets; empirically, in about 80% of scenarios the baked-in effects still hold perceptually. For the remaining large moves or lighting changes, yes, the intended remedy is to re-run BG_Gen once.

[1] ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation, CVPR 2025 [2] MRT: Masked Region Transformer for Layered Image Generation and Editing at Scale, CVPR 2026

We hope this addresses your concerns; please feel free to ask any further questions.

The rebuttal addressed most of my concerns. As a result, I decide to keep my original rating as weak accept.

Thank you for reviewing our rebuttal. We are glad that your concerns have been addressed, and we deeply appreciate your valuable feedback and support.

Comments on the paper

1 comment
Area_Chair_P5P3OpenReview ↗

Discussion phase will end on Nov 18

Dear Reviewers,

This is a kind reminder that discussion phase will end on Nov. 18. As the authors have provided the rebuttal, could you please check authors' rebuttal and update your thoughts? Your early feedback would make the discussion more smooth and informative.

Thank you.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.