Image Layers

• Image layers are structured visual components defined by properties such as transparency, color, and compositing order that facilitate effective image editing and analysis.
• Various methodologies use alpha blending, additive models, and neural network representations to decompose images into foregrounds, backgrounds, and effects.
• Layer-aware generative models and parsing techniques improve applications like recoloring, object isolation, and semantic retrieval by enabling local, controllable edits.

Searching arXiv for relevant papers on image layers, layered image decomposition, and layered image generation. Image layers are explicit or implicit strata used to represent visual content at different levels of editability, composition, or abstraction. In the literature, the term spans several related but distinct constructs: ordered RGBA elements whose alpha blending reconstructs a scene; additive color layers for recoloring; vector-like layers defined by transparency masks and colors; shared object prototypes composed with occlusion; and, in a broader neural-representation sense, intermediate CNN layers or repeatedly composed latent layers that encode different aspects of image content (Lin et al., 2017, Sbai et al., 2018, Monnier et al., 2021, Staszewski et al., 2020, Touvron et al., 2020). Across these formulations, the recurring goal is to separate what should be manipulated, analyzed, or retrieved independently—foreground objects, backgrounds, text, decorations, shadows, reflections, semantic categories, or appearance cues—while preserving a coherent reconstruction of the whole image.

1. Core meanings of image layers

The dominant usage treats an image as a composition of editable visual elements. In this sense, layers are not merely semantic regions: they usually carry appearance, transparency, and ordering. Text2Layer defines a two-layer image as a foreground image $F$, a background image $B$, and a mask $m$; LayerDecomp outputs a clean RGB background and a transparent RGBA foreground containing the target object plus associated visual effects; RevealLayer extends the idea to one background layer and multiple completed foreground RGBA layers; Referring Layer Decomposition defines the target as a complete prompt-specified RGBA layer recovered from a single RGB image (Zhang et al., 2023, Yang et al., 2024, Wang et al., 12 May 2026, Chen et al., 22 Feb 2026).

A second usage appears in recoloring and decomposition methods where layers are not object masks in the conventional sense. LayerBuilder decomposes an image or video into a small set of single-color additive layers, each with a layer color $c_j$ and a per-pixel contribution map $X_j$, so that pixels are reconstructed by linear combination. The paper explicitly states that these are not standard alpha-composited layers and not segmentation masks: pixels may belong to multiple layers simultaneously, layers are smooth and overlapping, and they represent color-source contributions rather than object ownership (Lin et al., 2017).

A third usage describes layer-like internal neural representations. For image retrieval, convolutional and fully connected CNN layers are treated as different image-description strata: fully connected activations emphasize semantic class, while convolutional or max-pooling layers preserve texture, color distribution, background, and scene structure. A more abstract extension appears in image-to-image translation, where a single latent residual block is repeatedly composed as $G(x)=Dec(f^n(Enc(x)))$; here the “power” of a layer becomes a controllable transformation parameter rather than an editable visual asset (Staszewski et al., 2020, Touvron et al., 2020).

A recurrent misconception is that image layers are equivalent to segmentation. Several papers reject that equation explicitly. CreatiParser distinguishes raster graphic design parsing from semantic segmentation by requiring category-appropriate editable outputs: text as a rendering protocol, stickers as transparent RGBA assets, and the remainder as a background layer. Referring Layer Decomposition likewise goes beyond visible masks by requiring complete RGBA layers, including occluded content when necessary (Chen et al., 21 Apr 2026, Chen et al., 22 Feb 2026).

2. Representation and compositing formalisms

The standard compositing model is alpha blending. Text2Layer writes the composed image as

$I \approx mF + (1-m)B,$

while LayeringDiff uses

$C = \alpha\cdot F + (1-\alpha)\cdot B.$

DreamLayer generalizes this to $k$ transparent layers and one global composed image,

$$I^{k+1} = \sum_{i=1}^k \left(I_\alpha^i \cdot I_c^i \cdot \prod_{f=i+1}^{k}(1-I_\alpha^f)\right),$$

and PSDiffusion formulates a multi-layer composition with one RGB background and multiple RGBA foregrounds, where each foreground contribution is attenuated by the layers in front of it (Zhang et al., 2023, Kang et al., 2 Jan 2025, Huang et al., 17 Mar 2025, Huang et al., 16 May 2025).

Other formulations replace alpha-over composition with additive or coordinate-based models. LayerBuilder reconstructs pixels as

$B$0

so each layer contributes additively according to a scalar mixing weight. “Unsupervised Image Decomposition in Vector Layers” instead defines each mask continuously by

$B$1

and composites one-color layers recursively as

$B$2

In “Authoring image decompositions with generative models,” intrinsic-style layers are composed in log space, beginning from

$B$3

and extended to include a distinct shading-detail layer (Lin et al., 2017, Sbai et al., 2018, Rock et al., 2016).

Explicit object-centric layer models introduce shared prototypes and occlusion. “Unsupervised Layered Image Decomposition into Object Prototypes” represents each layer as a transformed sprite with RGB appearance and transparency, then composes instantiated layers with an occlusion matrix $B$4. RevealLayer and LayerDecomp similarly make RGBA explicit, but target editable natural-image layers rather than prototype templates. This suggests a broad structural divide between formulations that treat layers as reusable symbolic or prototype units and formulations that treat them as image-conditioned RGBA assets (Monnier et al., 2021, Yang et al., 2024, Wang et al., 12 May 2026).

3. Decomposition and parsing of existing images

Earlier decomposition work often relied on strong priors over what a layer should look like. “Authoring image decompositions with generative models” learns separate conv-VAE or Laplacian conv-VAE priors for albedo, shading, and shading detail from proxy datasets such as Mondrian images, rendered 3D primitives, and material swatches, then optimizes latent codes so the decoded layers explain the input image in log space. “Unsupervised Layered Image Decomposition into Object Prototypes” instead learns a global dictionary of transparent sprites, together with per-layer transformations and an occlusion matrix, so images are reconstructed as compositions of transformed object prototypes. “Unsupervised Image Decomposition in Vector Layers” imposes a different prior: each layer has a single RGB color and a low-capacity coordinate-conditioned transparency mask, encouraging coherent editable regions rather than unrestricted texture synthesis (Rock et al., 2016, Monnier et al., 2021, Sbai et al., 2018).

Design-oriented decomposition methods shifted the focus from generic reconstruction to editability. LayerD treats raster graphic design decomposition as repeated extraction of current unoccluded top layers: a top-layer matting model predicts alpha, a background completion model reveals what lies behind, and inverse alpha blending recovers the foreground RGB. The paper adds a palette-based refinement step motivated by the assumption that graphic-design layers often exhibit uniform appearance. CreatiParser addresses a related but more heterogeneous setting by decomposing a raster design into three layer types—text, sticker, and background—using a vision-LLM for text rendering protocols and a multi-branch diffusion architecture with RGBA sticker generation. Text is thus recovered as renderable instructions rather than as pixels (Suzuki et al., 29 Sep 2025, Chen et al., 21 Apr 2026).

Natural-image decomposition increasingly emphasizes effects, occlusion, and promptability. LayerDecomp outputs a clean RGB background and a transparent RGBA foreground that retains object-associated shadows and reflections, using a consistency loss to learn such effects even when real images lack ground-truth foreground layers. RevealLayer decomposes an RGB image into a background and multiple RGBA foregrounds under user-specified boxes, with Region-Aware Attention and an Occlusion-Guided Adapter to separate visible and hidden content. Referring Layer Decomposition generalizes the task further: given an RGB image and a prompt such as a point, box, mask, text description, or combination, the model predicts a complete RGBA layer rather than a visible mask fragment (Yang et al., 2024, Wang et al., 12 May 2026, Chen et al., 22 Feb 2026).

These methods collectively challenge the idea that decomposition is only about visible partitioning. In several papers, the target layer must be complete, editable, and compositional, which may require amodal completion, explicit alpha recovery, background cleaning, or symbolic reconstruction of text. A plausible implication is that layer decomposition has become less a special case of segmentation and more a structured inverse graphics or inverse design problem.

4. Layer-aware generative models

Layered generation moved from two-layer foreground/background models to joint multi-layer synthesis. Text2Layer trains a composition-aware latent diffusion model that generates foreground, background, layer mask, and composed image simultaneously from text. LayerDiff extends this to one background layer, multiple foreground layers, and per-foreground masks, using inter-layer attention, text-guided intra-layer attention, a layer-specific prompt-enhanced module, and self-mask guidance. Both papers argue that layered outputs provide better control than flat image generation followed by post hoc extraction (Zhang et al., 2023, Huang et al., 2024).

Subsequent work diverged into two strategies. One strategy generates layers directly and jointly. DreamLayer models one background layer, $B$5 transparent foreground layers, and an additional global composed image, and introduces Context-Aware Cross-Attention, Layer-Shared Self-Attention, and Information Retained Harmonization to enforce coherent occlusion, layout, and latent-space fusion. PSDiffusion likewise generates one RGB background and multiple RGBA foregrounds in a single diffusion run, using layer cross-attention reweighting for layout alignment and partial joint self-attention for appearance alignment. LASAGNA jointly generates a photorealistic composite image, a clean background, and a transparent foreground with visual effects, with conditioning on text, masks, background images, or foreground images for real-world editing workflows (Huang et al., 17 Mar 2025, Huang et al., 16 May 2025, Yang et al., 21 Jan 2026).

The second strategy is generation followed by disassembly. LayeringDiff first synthesizes a composite image with an off-the-shelf model such as SDXL, then estimates an alpha mask and decomposes the image into foreground and background using a diffusion-based decomposition module plus high-frequency alignment networks. The paper’s argument is that large-scale layered training data are hard to obtain, whereas high-quality composite generation already exists. This suggests a trade-off between direct layered modeling and leveraging pretrained whole-image generative priors (Kang et al., 2 Jan 2025).

Graphic-design generation introduces additional structural constraints. LaDe represents a design as one full composed image plus a flexible number of semantically meaningful RGBA layers, allows a layer to contain multiple disconnected regions, and uses an LLM-based prompt expander together with a latent diffusion transformer and 4D RoPE over spatial position, layer index, and token role. In that formulation, text-to-image is the special case $B$6, text-to-layers jointly generates the full design and its layers, and image-to-layers freezes the input design as a condition while denoising only the layer latents (Lungu-Stan et al., 18 Mar 2026).

5. Editing, retrieval, and layer-conditioned computation

Layered representations are attractive largely because they make edits local and reusable. LayerBuilder is explicit: once the contribution maps $B$7 have been computed, recoloring is reduced to replacing layer colors $B$8 and re-summing the layers. The same representation supports automatic recoloring suggestion, texture synthesis, and color-based filtering. LayerDecomp uses its clean background plus transparent foreground with visual effects for object removal, object movement, resizing, recoloring, and sequential extraction of multiple objects into multiple layers. LASAGNA evaluates three editing paradigms—direct instruct editing, explicit layer editing, and layer editing with visual effects—and reports that adding effect-aware layers yields the best CLIP-FID/FID across recolor, movement, and combined edits (Lin et al., 2017, Yang et al., 2024, Yang et al., 21 Jan 2026).

Direct editing of already layered assets motivates a different formulation. LimeCross is a training-free method that edits one selected RGBA layer according to text while leaving the remaining layers unchanged. It composes all non-target layers into an opaque context scene, encodes target and context into separate token streams, and lets the target stream attend to context through a bi-stream attention mechanism. Only the target stream is updated and decoded, and a late target-only refinement stage suppresses context-induced alpha corruption. The paper’s central principle is effectively to read context without collapsing the layered representation (Morita et al., 11 May 2026).

Graphic-design parsing extends editability beyond pixels. CreatiParser represents text as a rendering protocol containing geometric, semantic, appearance, and relational attributes; sticker elements are produced as transparent RGBA assets; and the remainder becomes a background layer. This means replacing text need not involve bitmap editing at all. A plausible implication is that “layer” increasingly denotes the representation most natural for the element type, not necessarily a raster image for every component (Chen et al., 21 Apr 2026).

A broader but related line of work uses neural-network layers as image descriptors or transformation controls. In image retrieval, the descriptor

$B$9

combines fully connected activations with aggregated convolutional features, and on IMAGENET1M the combined descriptor achieved the best average normalized histogram distance, $m$0, relative to FC-only, convolution-only, and IMAGENET1M-provided baselines. In image-to-image translation, repeated application of a shared latent residual block supports controllable transformation strength and adaptive stopping; on Gaussian denoising with standard deviation 30, discriminator-based adaptive test-time selection improved PSNR by $m$1 dB relative to a fixed test-time number of iterations (Staszewski et al., 2020, Touvron et al., 2020).

6. Datasets, evaluation, and recurring limitations

As image layers became a first-class target, evaluation shifted from flat-image realism to layer-centric metrics. Text2Layer built a filtered layered dataset of about 57.02M samples from LAION-Aesthetics, with 57M for training and 20K for testing. LayerDiff introduced MLCID with about 2M training examples and a 27k-sample test set. DreamLayer constructed a 408,187-sample multi-layer dataset with 305,801 two-layer, 87,571 three-layer, and 14,815 four-layer images. PSDiffusion introduced Inter-Layer with 30,000 multi-layer images, each containing 3–6 layers and artist-grade alpha mattes. LASAGNA-48K contains over 48K triplets of composite image, clean background, and foreground with visual effects. RevealLayer-100K provides 100K natural-image tuples plus a 200-image RevealLayerBench, while RefLade contains about 1.11M image-layer-prompt triplets, including 100K manually curated high-fidelity layers (Zhang et al., 2023, Huang et al., 2024, Huang et al., 17 Mar 2025, Huang et al., 16 May 2025, Yang et al., 21 Jan 2026, Wang et al., 12 May 2026, Chen et al., 22 Feb 2026).

The metrics likewise diversified. Graphic-design parsing uses layer IoU for text and sticker masks, font accuracy, attribute accuracy, and per-layer RGB $m$2; CreatiParser-RL reports $m$3-IoU $m$4, $m$5-IoU $m$6, Font $m$7, Attr $m$8, and average RGB $m$9 $c_j$0 on Parser-40K. RevealLayerBench evaluates background PSNR, SSIM, LPIPS, FID and foreground PSNR, LPIPS, FID, SoftIoU; RevealLayer reports background PSNR $c_j$1, foreground PSNR $c_j$2, and foreground SoftIoU $c_j$3. Referring Layer Decomposition introduces a Human Preference Aligned score combining visible-content preservation, completion, and compositional faithfulness, and reports Pearson $c_j$4 correlation with human Elo. LimeCross adds mFID-$c_j$5 to assess alpha-channel stability in layered editing workflows (Chen et al., 21 Apr 2026, Wang et al., 12 May 2026, Chen et al., 22 Feb 2026, Morita et al., 11 May 2026).

Several limitations recur across the literature. Layer granularity remains ill-posed and partly subjective in graphic design decomposition. LayerBuilder acknowledges failures when semantically different but similarly colored regions are hard to separate automatically. LayerDecomp explicitly focuses on common effects such as shadows and reflections and leaves richer phenomena such as smoke and mist for future work. LASAGNA currently focuses on single-object layer generation. RevealLayer notes degradation under severely inaccurate boxes, heavy occlusion, transparent or translucent objects, and dense repetitive textures. RefLade states that shadows, reflections, rain, dust, and similar factors are ignored in its current benchmark. These constraints indicate that, despite rapid progress, image layers are still easiest to define when compositing order, effect semantics, or content type are relatively controlled (Suzuki et al., 29 Sep 2025, Lin et al., 2017, Yang et al., 2024, Yang et al., 21 Jan 2026, Wang et al., 12 May 2026, Chen et al., 22 Feb 2026).

Taken together, the literature presents image layers as a unifying representation for editable graphics, structured scene decomposition, multi-layer generation, and even neural-image analysis. The common thread is not a single compositing equation or architecture, but a design principle: meaningful image structure is easier to generate, manipulate, and evaluate when it is separated into components whose interactions are explicit rather than flattened away.