acceptodds
Withdrawn from ICLR 2027

PrismLayers: Open Data for High-Quality Multi-Layer Transparent Image Generative Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-quality multi-layer transparent image generation enables layer-level editing, with graphic design as its most crucial testbed. Progress on both fronts lags behind text-to-image generation, as web data is flattened, professional source files are scarce, and alpha supervision is costly to annotate. To break this bottleneck, we build an end-to-end data route that turns existing single-layer generators into a coherent supervision source for both layered synthesis and photorealistic object-transparency editing. We release the first large-scale series of open, ultra-high-fidelity multi-layer datasets—PrismLayers, PrismLayersPlus, PrismLayersPro, and PrismLayersReal—comprising 200K, 100K, 20K, and 1K multi-layer transparent images and accurate alpha mattes, respectively, and further construct a photorealistic object-transparency editing corpus, PrismLayersEdit—containing 80K paired edit targets across 10 opacity levels. On top of this, we introduce a systematic training-free agentic pipeline that orchestrates pre-trained diffusion models into coherent layered generation rather than isolated layer prediction. Its core is a two-stage design: LayerFLUX produces clean, high-detail transparent layers with reliable alpha quality, while MultiLayerFLUX composes them into semantically aligned full designs under human-annotated layout guidance. A strict quality-control loop (artifact filtering, semantic consistency checks, and human selection) further pushes visual fidelity to production-level quality, and an advanced Qwen-Image-Edit-based engine extends the pipeline to construct paired supervision for real-image object-transparency editing. Finally, fine-tuning the state-of-the-art ART model on PrismLayersPro yields ART+, which beats the original ART in 60% of head-to-head user comparisons. Our results establish a strong foundation for editable, precise, and aesthetically competitive multi-layer transparent image generation and introduce object transparency editing as a new frontier for future research.

Primary area:
datasets and benchmarks
Keywords:
Multi-layer Image Generation, Transparent Image Dataset, Data-centric Generation Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pze7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the shortage of high-quality multi-layer transparent image datasets, which has limited the development of editable image generation models. The authors introduce PrismLayers, a family of synthetic multi-layer transparent image datasets, together with an automated data construction pipeline.

Critical strengths
  1. High-quality multi-layer transparent image generation is an important capability for editable visual content creation. Existing image generation models primarily produce flattened RGB images, which are difficult to manipulate at the object or layer level. The paper correctly identifies the lack of high-quality transparent training data as an important limitation and proposes a practical approach to generating such supervision at scale.
  2. The proposed data generation pipeline is practical and scalable. The combination of pretrained image generators, background removal, semantic layout guidance, and automatic quality filtering provides a relatively straightforward way to construct large-scale multi-layer datasets.
Critical weaknesses
  1. The novelty of the proposed generation framework is relatively limited, and the effectiveness of its individual components is insufficiently analyzed. The proposed LayerFLUX framework primarily combines two existing techniques: generating images on a uniform background using a prompt suffix and applying a pretrained background-removal model to extract the foreground. Although this integration is practically useful, the core technical contributions appear relatively incremental compared with prior approaches such as MAGICK, LayerDiffuse, and existing layout-guided image generation pipelines.
  2. More importantly, the paper does not sufficiently demonstrate which components are responsible for the improvements. For instance, the dataset construction pipeline contains several potentially important operations. However, the paper lacks a systematic ablation study that isolates the contribution of these components to downstream generation performance.
  3. The paper introduces FLUX-Multi-Layer-Bench to evaluate multi-layer generation against high-quality images generated by FLUX.1-dev. While this benchmark is useful for evaluating similarity to modern text-to-image outputs, it introduces an important potential bias. Specifically, the benchmark references are generated using FLUX.1-dev, which is also the primary generator used to construct PrismLayers. Therefore, improvements on this benchmark may partly reflect better alignment with the distribution of the data generation model rather than general improvements in multi-layer generation.
Questions for authors

Refer to weakness

Reviewer_EWx6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper (1) proposes a training-free pipeline for synthetic multi-layer transparent images and (2) releases a family of datasets built with the pipeline. The layouts are from 800K crawled graphic designs. Then, a training-free module (LayerFLUX) generates each layer with a text-to-image model (FLUX.1-dev) and mattes it out. The next module (MultiLayerFLUX) stacks the layers according to the layout. An artifact classifier and a layer scorer (TIPS) filter the results into PrismLayers (200K) and, after restyling, PrismLayersPlus (100K). Adding manual selection gives PrismLayersPro (20K). They also build a 1K photorealistic set and an object-transparency editing corpus. In a user study, 58% favor ART trained on their data. ART+ also gets better FID and TIPS on a new benchmark.

Critical strengths

Existing open sets are either small or low quality, while the large ones are closed, so the lack of open multi-layer data is the main bottleneck for multi-layer transparent image generation. The paper fills this gap with a training-free pipeline that synthesizes new data and can be rerun with other generators, such as Qwen-Image, as they improve. The user study also favors ART+(20k scratch), supporting the claim that open synthetic data can train a multi-layer model without closed data.

Critical weaknesses

My main concern is whether the metrics measure the quality of the data. (1) TIPS is used to filter the data (L155-156, L278-280) and then used to show that the data and the models trained on it are better (Figure 2b, Table 1). This is circular, so it is unclear whether TIPS measures quality. (2) FID measures how close the distribution of the outputs is to that of the reference images. ART has the best FID on DESIGN-MULTI-LAYER-BENCH, which is attributed to similarity to ART’s training data (L457-458). By the same logic, FLUX-MULTI-LAYER-BENCH may favor ART+ because its references are images from FLUX.1-dev (L657-659) and ART+ is trained on FLUX.1-dev outputs. Overall, apart from the user study, the numbers do not show the data quality.

Two more claims are not supported. ART+(100k) and ART+(20k) start from MLTD (Table 10), but ART+(100k) is worse than ART+(20k) in all of Table 1, which does not support the claim that scaling works (L430-431). The editing corpus is listed as a contribution (L047-048), but training and evaluating editors on it is in future work (L483-485). It is also unclear which version the claim about ART+ refers to, since the abstract describes fine-tuning ART on PrismLayersPro (L056-058), but the user study tests ART+(20k scratch) (L425-426).

Questions for authors
  1. Can you report how well TIPS agrees with human pairwise preferences?
  2. Can you report the user study on DESIGN-MULTI-LAYER-BENCH? Human judgment does not depend on reference images, so it could settle the FID question.
  3. Which version does the claim in L56-58 refer to? Can you report the user study for ART+(20k)?
  4. ART+(100k) is worse than ART+(20k) on every metric in Table 1. Can you show results where the filtered data from the pipeline improves the model to support the claim in L430-431?
  5. Can you show results of fine-tuning with the editing corpus?
Reviewer_xBCHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PrismLayers, a synthetic dataset for multi-layer transparent image generation. Its pipeline extracts layouts and layer descriptions from existing layered designs, generates individual elements with FLUX, obtains alpha mattes using an off-the-shelf background-removal model, and composes the resulting layers according to the extracted layouts. Quality control combines an artifact classifier, the proposed Transparent Image Preference Score (TIPS), and human selection. The datasets include PrismLayers (200K), Plus (100K), Pro (20K), and Real (1K). The paper also introduces a separate data engine for object-transparency editing using Qwen-Image-Edit and latent interpolation.

Validation includes comparisons of single-layer generation and matting methods, dataset quality analyses, and downstream training of ART variants with PrismLayers data. The resulting ART+ achieves approximately 58–60% preference over ART in a small user study and improves quantitative results on the FLUX-based benchmark, although improvements are not consistent across all benchmarks. Downstream training and evaluation using the transparency-editing corpus are left for future work.

Critical strengths

A useful data contribution for editable image generation. PrismLayers addresses the scarcity of high-quality multi-layer training data by providing synthetic designs with individual RGBA layers and associated captions. Its combination of scale, diverse styles, and complex layered compositions could support research beyond flattened image generation. Evidence of downstream utility beyond dataset aesthetics. The paper evaluates its data by training ART+ and comparing it with ART, rather than relying solely on visual examples or dataset quality scores. ART+ receives approximately 58–60% preference across the reported human-evaluation dimensions and improves FID on the FLUX-based benchmark. Although these gains are not universal across benchmarks, they provide concrete evidence that the proposed data can benefit an existing multi-layer generation model.

Critical weaknesses

The evaluation does not sufficiently establish transparent-layer quality independently of aesthetic preference. TIPS is trained from aggregated RGB-oriented scoring models and is used both to filter the dataset and to evaluate its quality, creating a risk of favorable selection under the same metric. Its training assumes that alpha quality is already satisfactory, while the paper does not establish its agreement with independent human judgments of transparency defects. Layer visualizations are useful, but do not resolve this limitation. Since high-quality transparent layers are the central contribution, stronger validation of alpha accuracy, boundary quality, and editability is necessary.

The downstream evidence does not clearly isolate the benefit of the proposed data. The FLUX-based benchmark uses references generated by the same model family that produces the training data, making improved FID difficult to separate from distribution alignment. Performance on the existing design benchmark is mixed. The human study provides complementary evidence, but covers only 40 samples and reports no confidence intervals. A controlled comparison with matched training budgets, independently sampled test cases, explicit input conditions, and style-stratified evaluation would more convincingly establish that the dataset improves general multi-layer generation rather than primarily changing its visual style.

Questions for authors

Advantage over existing datasets: What evidence isolates the benefit of PrismLayers relative to existing layered datasets, particularly MLTD and Crello, under comparable model architectures, training budgets, and data quantities? Please clarify which existing experiments support this comparison and how the qualitative quality ratings in Table 9 were determined.

Validity of TIPS: Has TIPS been evaluated against independent human judgments of transparent-layer quality, particularly alpha-boundary defects, background leakage, and incorrect semi-transparency? Since TIPS provides both filtering and evaluation, how are its training data, dataset-selection pool, and evaluation samples separated?

Evaluation comparability: For Figure 8 and the user study, did all methods receive equivalent global prompts, layer descriptions, style instructions, and layout information? Please provide the exact input conditions and explain how the 40 examples were selected. Were the reported preferences statistically reliable when accounting for variation across both examples and participants?

Release scope and data independence: Which images, alpha mattes, layouts, captions, and filtering models will be released, and under what usage terms given the commercial-source designs? Please also clarify overlap among PrismLayers, Plus, and Pro, whether train/test separation occurs at the original-design level, and whether the transparency-editing corpus contains 50K or 80K pairs.

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

Related papers

Discussion (0)

Sign in to comment.