# UniWorld-Design：从像素生成到图层原生设计

- 来源：HuggingFace Daily Papers（社区热门论文）
- 发布时间：2026-08-04 08:00
- AIHOT 分数：57
- AIHOT 链接：https://aihot.virxact.com/items/cmsfhv0y81tjrro2exrywjmkr
- 原文链接：https://arxiv.org/abs/2608.03971

## AI 摘要

UniWorld-Design 将图像生成从平面像素合成重构为结构化视觉组合，以语义 RGBA 图层作为生成、理解和编辑的原子单元。框架包含两个模型：Text-to-RGBA（T2RGBA）直接从文本生成独立 RGBA 素材，Image-to-Layer（I2L）根据图像、全局指令和逐层提示联合生成有序的完整语义 RGBA 图层。

## 正文

Abstract

We introduce UniWorld-Design, a framework that redefines image generation from flat pixel synthesis to structured visual composition, with semantic RGBA layers as the atomic units of generation, understanding, and editing. Our key insight is that pixels define how an image is rendered, whereas layers define how an image is created, understood, and edited. Just as human designers create and manipulate visual content through layers rather than raw pixels, UniWorld-Design equips multimodal generative models with a layer-native design space. UniWorld-Design comprises two models. The Text-to-RGBA (T2RGBA) model generates standalone RGBA assets directly from text. The Image-to-Layer (I2L) model conditions on a finished image, a global instruction and per-layer prompts, and jointly produces ordered, complete semantic RGBA layers. Its instruction interface supports top-level decomposition, recursive decomposition and targeted extraction, making layering an instruction-addressable operation for agentic editing. Because I2L learns complete semantic objects rather than visible-pixel partitions, its layers stay usable when moved or removed. On the Crello benchmark, I2L reduces per-layer RGB L1 error by and achieves a relative improvement in Alpha Soft IoU over Qwen-Image-Layered. Separately, T2RGBA achieves the highest CLIP Score, outperforming LayerDiffuse and OmniAlpha.

[ Project Website]https://rabbitvis.rabbitpre.com/blog ††∗Full author list in Contributions.

媒体内容 · 前往原文查看

Figure 1: UniWorld-Design generates and decomposes design assets natively in RGBA layers.(a) T2RGBA: one prompt, one standalone transparent asset. (b) I2L follows structured prompts for instruction-controlled semantic decomposition; any returned layer can be supplied to a subsequent call for recursive decomposition.

1 Introduction

Text-to-image models have made remarkable progress in producing visually compelling images from language [esser2024sd3, lin2025uniworld, diao2026sensenova, Yan_2026_CVPR], but they still largely synthesize a flat RGB canvas. Pixels determine how an image is rendered; they do not explicitly describe how it is composed, which semantic elements it contains, or how those elements should be manipulated. Human designers instead build posters, illustrations and social media graphics as ordered layers—backgrounds, subjects, decorations and text—that can be independently selected, moved, recolored, replaced or removed [yamaguchi2021canvasvae, suzuki2025layerd]. The limitation of a flat output is therefore not only that editing is difficult, but that editable structure is absent from the output space itself.

Once a layered design is flattened into a raster image, its editable structure is lost: individual objects, their layer order and hierarchy, and content hidden by occlusion are no longer stored explicitly. Any subsequent attempt to edit a person, a text block or the background must first reconstruct that structure from pixels, and this recovery is generally incomplete. Whole-image instruction editing may alter content that should remain unchanged [brooks2023instructpix2pix, yin2025qwenimagelayered]; masks can localize edits [couairon2023diffedit]; and foreground segmentation can extract visible regions [zheng2024birefnet, qin2022dis]. However, these methods do not reconstruct the original layer stack or content hidden by occlusion [zhu2017semanticamodal, ozguroglu2024pix2gestalt]. The limitation therefore lies not only in segmentation accuracy, but also in the absence of editable structure from the raster representation itself. PSD-like workflows provide the natural reference: each element is stored from the outset as an independently editable object with an explicit place in the composition.

Recent image-to-layer models, including Qwen-Image-Layered and OmniPSD, demonstrate that editable layer structures can be recovered from finished images [yin2025qwenimagelayered, liu2025omnipsd]. Agentic editors likewise show the value of planning and structured execution [wang2024genartist, ye2026agentbanana, yun2026redesign]. However, a model-level question remains: can generation and decomposition operate directly in a layer-native visual space, with language controlling what should be separated, at what granularity, and with which semantic role? In such a space, layers rather than raw pixels become the atomic units of generation, understanding and editing.

We propose UniWorld-Design, a framework that represents images as ordered sets of semantic RGBA layers. Each layer is a complete, independently addressable visual object that can be generated, decomposed, extracted, edited, moved, deleted and recomposed. The framework comprises two models. The Text-to-RGBA (T2RGBA) model generates reusable RGBA assets from text. The Image-to-Layer (I2L) model takes a finished image, a global task instruction and per-layer prompts, and jointly produces an ordered stack of complete semantic RGBA layers on a shared canvas. Together, they connect layer generation and layer decomposition within one representation, allowing visual content to move between language, finished images and editable structure. Figure 1 summarizes these two capabilities.

Unlike a fixed segmentation taxonomy, I2L exposes decomposition as an instruction-addressable operation. Top-level decomposition separates a complete image into specified semantic layers; recursive decomposition refines a selected layer; and targeted extraction separates the requested content from the unselected remainder. An external agent can therefore decide what should become an independent layer, adjust the granularity of an existing decomposition, or extract a target before editing and recomposition. UniWorld-Design provides reusable, language-addressable object states as an interface for external planning systems.

Three design principles underpin this formulation. First, both models treat RGBA as a native compositional representation rather than merely an output format or post-processing result. Encoding each object with its own color and alpha makes it independently placeable and directly composable. Second, I2L preserves the finished image as a shared visual condition while binding each ordered output to its corresponding semantic prompt, supporting joint consistency without losing independent control. Third, it learns complete semantic objects rather than visible-pixel partitions. Visible partitions may recompose the input accurately yet reveal holes or truncated regions when an occluder is moved; complete layers retain the hidden content needed for subsequent editing. This establishes a closed loop in which T2RGBA generates semantic assets, I2L converts finished designs back into language-addressable layers, and selected layers can be refined, modified and recomposed. Figure 2 illustrates how these operations can serve an external design agent.

Figure 2: From pixel generation to agentic layer-native design. A pixel-only interface exposes a monolithic raster, requiring object structure to be inferred again. UniWorld-Design instead exposes a persistent ordered state of semantic RGBA layers. An external design agent can invoke T2RGBA generation and I2L decomposition, recursively refine a returned layer, edit selected layers with other tools, and recompose the state.

Contributions.

This paper makes three contributions:

•

A layer-native formulation that treats semantic RGBA layers as the atomic unit of generation, decomposition and editing. By exposing persistent, independently manipulable object states, it provides a structured interface over which external design agents can plan and execute generation, decomposition, refinement, editing and recomposition.

•

Two models that realize this formulation: T2RGBA generates reusable RGBA assets, while I2L provides an instruction-addressable interface for decomposing finished images into ordered, complete semantic layers at controllable granularity.

•

A systematic evaluation covering per-layer fidelity, transparency, editability and RGBA asset generation.

Compared with Qwen-Image-Layered, I2L reduces per-layer RGB L1 by , improves Alpha Soft IoU by , reduces blank-layer generation by , and achieves a VLM score of versus . T2RGBA reaches the highest CLIP Score among the models compared (). Alpha boundaries and dense text remain challenging.

2 Related Work

RGBA generation.

Transparency-aware generation has followed two broad strategies. One preserves a pretrained RGB representation and attaches alpha through latent offsets, auxiliary branches, attention cues or separate decoders [zhang2024layerdiffuse, pu2025art]. The other learns a native RGBA representation for reconstruction, layered generation or multi-task generation [wang2025alphavae, liu2025omnipsd, yu2025omnialpha]. Our framework follows the latter direction but assigns RGBA two roles across two models: T2RGBA generates standalone RGBA assets, whereas I2L represents multiple ordered semantic layers.

Layer decomposition.

Qwen-Image-Layered jointly generates semantically disentangled RGBA layers and supports recursive re-decomposition [yin2025qwenimagelayered]; OmniPSD targets text-to-PSD generation and image-to-PSD recovery [liu2025omnipsd]; and LayerD sequentially extracts elements from raster graphic designs [suzuki2025layerd]. Related work also provides layered training data and editability-oriented evaluation protocols [chen2025prismlayers, tudosiu2024mulan, rowles2026stablelayers]. Qwen-Image-Layered does not expose per-layer semantic prompts or targeted extraction controlled by text instructions. In contrast, I2L supports instruction-controlled semantic decomposition by binding a global task instruction and per-layer prompts to ordered outputs, unifying top-level decomposition, recursive decomposition and targeted extraction in one interface. It targets complete semantic layers rather than visible-pixel partitions.

Agentic image editing.

GenArtist uses a multimodal planner to decompose requests, select visual tools, and verify and revise intermediate results [wang2024genartist]. More recent systems introduce layer-aware execution. Agent Banana isolates a mask-localized high-resolution crop, edits it, and fuses it back, so its “layer” is a local execution patch [ye2026agentbanana]. ReDesign instead grows an editable JSON hierarchy by selecting and verifying heterogeneous extraction tools, including a layered decomposition model [yun2026redesign]. These works demonstrate the value of structure in agentic editing, but focus on planner–executor systems, local patch preservation or multi-tool hierarchy recovery. In contrast, our native RGBA decomposition model exposes complete semantic layers as persistent, language-addressable object states, providing external design agents with a structured interface for planning and executing edits directly in a layer-native visual space.

3 Method

3.1 Preliminaries

Flow matching.

Let be a clean sample from the data distribution with a corresponding condition (text for T2RGBA; text and a composite latent for I2L), and let be Gaussian noise. Rectified flow [liu2023rectifiedflow, lipman2023flowmatching] interpolates between them as

(1)

where is clean data and is noise. The model learns the velocity by mean-squared regression [esser2024sd3], and inference integrates the learned field from noise to data. All variables below denote autoencoder latents. We use for a layer index, for the number of target layers, for layer , and for the composite-image condition.

DiffusionNFT.

DiffusionNFT [zheng2025diffusionnft, li2025uniworld] performs reward-driven post-training on outputs sampled under the same condition . Let be candidate ’s group-normalized optimality score, its corresponding target velocity and the velocity predicted by the EMA-updated data-collection policy. With the DiffusionNFT implicit-policy parameterization and , the core objective is

(2)

High pulls the model toward the sampled target through ; low pushes it away through the reflected . Both branches are parameterized by the same flow field, so the sampler is unchanged. Reward normalization, adaptive loss weighting and EMA policy updates follow DiffusionNFT [zheng2025diffusionnft].

3.2 RGBA Autoencoder with RGB Latent Alignment

Channel extension.

Following recent RGBA autoencoders [wang2025alphavae, yin2025qwenimagelayered, liu2025omnipsd], we extend only the encoder’s first and decoder’s last convolution from three to four channels. The pretrained RGB weights are copied, both new alpha filters are zero-initialized, and the decoder’s alpha bias is set to one. The extended model therefore starts as the original RGB autoencoder with an opaque alpha channel.

Training.

We train on a mixture of transparent RGBA images and opaque RGB images (with alpha set to one). The objective combines RGB and alpha reconstruction, perceptual, KL, adversarial and RGB-latent alignment terms:

(3)

Here is the optimization step, is the adversarial-loss start step and is the adaptive gradient-norm ratio. The RGBA encoder receives an opaque alpha channel for an RGB sample , while the frozen teacher receives its three RGB channels. Applying this alignment only to samples that originate as RGB encourages their encodings to remain close to the pretrained latent space without directly applying the teacher term to transparent examples. Autoencoder training uses straight alpha to retain color beneath transparency, whereas diffusion training represents RGB after compositing over white.

3.3 Text-to-RGBA Generation

Training data is an internally curated dataset of text paired with transparent RGBA images, including design assets, cutout subjects and vector-derived illustrations. Transparent RGBA images constitute the majority of the training mixture, while opaque images are also included. Fine-tuning starts at -equivalent area and then continues at .

Both T2RGBA and I2L use the two-stage training pipeline detailed in Section 3.6: progressive distillation followed by post-training with DiffusionNFT [zheng2025diffusionnft]. Their task-specific rewards are specified in Equation 6. The layering model of Section 3.4 starts from the post-trained T2RGBA model rather than from the base model. The T2RGBA results in Section 4.3 are those of the model obtained after distillation and DiffusionNFT post-training, sampled at -equivalent area.

3.4 Instruction-Controlled Image-to-Layer Generation

Figure 3: LIB-MMDiT overview.Left: the attention mask broadcasts the global instruction to all image queries while restricting each layer prompt to its corresponding layer. Right: LIB-MMDiT jointly generates aligned RGBA layers from an input image and structured layer instructions.

I2L maps a composite image and an instruction to ordered, complete semantic RGBA layers on the same canvas. Each target represents a requested semantic role or, for targeted extraction, the unselected remainder; occluded content is retained. The model is trained so that alpha-compositing the layers from back to front approximates the input. Instruction-controlled decomposition creates two correspondence problems. First, each semantic description must control the intended output layer rather than the whole stack. Second, tokens from different layers must remain aligned to the same canvas while retaining distinct layer identities and order. We call the resulting architecture Layer–Instruction Binding MMDiT (LIB-MMDiT). It combines layer–instruction binding attention with layer-indexed rotary positions to address the two correspondence problems, respectively (Figure 3).

Layer–instruction binding attention.

We assign tag to the global instruction and composite condition, and tag to the prompt and image tokens of target layer . An image query reads global text and its matching prompt, but not prompts assigned to other layers. Image-image attention remains unrestricted, allowing the condition and all targets to jointly resolve occlusion and stacking. Text-text attention is also unrestricted; the binding mask applies specifically when an image query attends to a text key. The global task instruction is therefore broadcast to the stack, while each layer description is bound only to its intended output.

Layer-indexed 3D rotary positions.

The base transformer assigns every image token three rotary coordinates , with a constant first coordinate and spatial row and column. Following the layer-indexed positioning of Qwen-Image-Layered [yin2025qwenimagelayered], we reuse that constant coordinate for layer identity rather than adding another positional axis. Let denote the rotary coordinates of an image token, and let denote the visibility of a text key to an image query , with and the image- and text-token sets. The composite uses layer index , and target layer uses index , with the bottom-most layer. Equal values align the same canvas location across the stack, while the first coordinate distinguishes layer order. Text keeps the base model’s position scheme; its layer binding is handled by . This reuse introduces no new RoPE parameters and preserves the base position encoding for a single target at layer zero.

Together, the two mechanisms are summarized as

(4)

Conditioning and training.

The global instruction, per-layer prompts, encoded composite and target-layer latents enter one LIB-MMDiT sequence. The composite is kept clean at , re-injected at every denoising step and excluded from the loss; only the target layers are noised and supervised, using one shared timestep per sample. The layer-binding and layer-indexing mechanisms are retained throughout fine-tuning, distillation and post-training. The exact image-token and text-prompt packing order is specified in Appendix C.

3.5 Building Semantic Layer Trees from PSD Documents

Figure 4: From one PSD document to three training tasks.(1) PSD layers are grouped into a semantic layer tree. (2) Complete semantic layers retain content behind occluders, unlike visible-pixel partitions. (3) The tree yields top-level decomposition, recursive decomposition and targeted extraction instances.

Training I2L requires layer stacks whose layers remain usable when moved or removed. Synthesizing such stacks from generated assets is circular: their layer statistics inherit the generator’s biases, and any occluded content must itself be generated. We instead use designer-authored PSD documents, where layer content hidden in the final composite remains available in the source file. We render the PSD layers, merge related authoring units into semantic objects and organize them into a hierarchy. A vision-language model assists only with grouping; all target pixels come from the original PSD layers. Figure 4 summarizes the construction and its three derived tasks.

Preserving occluded content.

When constructing a target layer, we preserve its source pixels even where they are hidden by layers above it. Moving or removing an occluding layer therefore reveals the underlying content rather than a transparent hole. Visible-only cutouts are not used as training targets.

Tree-derived supervision.

Each tree node defines a composite and its semantic children. The root provides top-level decomposition examples, internal nodes provide recursive decomposition examples, and selected nodes together with the unselected remainder provide targeted extraction examples. The tree recomposition is used as the image condition so that it is consistent with the retained targets. The three task types are mixed during fine-tuning.

3.6 Progressive Distillation and Post-Training

Emitting several full-resolution layers in one sequence multiplies its length, so sampling cost dominates inference. After supervised fine-tuning (Section 3.4), we use progressive distillation to reduce sampling cost and post-training with DiffusionNFT to optimize layer-level output quality. The post-training adapter is merged to obtain the final layering model.

Empirical choice of trajectory supervision.

We also experimented with distribution-matching distillation [yin2024dmd, yin2024dmd2]. In our runs on layered RGBA generation, we observed alpha predictions tending toward full opacity, misalignment between RGB content and alpha masks, and weaker condition following and cross-layer consistency. We therefore adopt progressive adversarial distillation.

Progressive adversarial distillation.

Following progressive distillation [salimans2022progressive] and the progressive adversarial formulation of SDXL-Lightning [lin2024sdxllightning], we train the student to predict a direction that reaches the flow location obtained after multiple teacher steps. Specifically, we sample an interval from the coarser student schedule, where under the convention of Equation 1, and add noise to real target latents to obtain the starting state . The frozen teacher then takes Euler steps, while the student moves directly to the same time in one step:

(5)

Here denotes the teacher’s multi-step update. The clean condition layer and text form ; the condition layer is never noised and is excluded from the sum over target layers. The endpoint term anchors the student prediction to the teacher’s next flow location. In alternating adversarial updates, a discriminator initialized from the transformer backbone treats the teacher endpoint as real and the student endpoint as fake at the same target timestep and under the same conditioning. This term discourages the over-smoothed endpoints produced by distance regression alone. Distillation yields an eight-step student, which remains the evaluation schedule after post-training.

Task-specific rewards.

For DiffusionNFT post-training, the two models use different reward pipelines. For T2RGBA, every decoded RGBA sample is evaluated by three reward heads. Let denote the alpha-following score, which reads the generated alpha channel and the prompt’s transparent/opaque label, rewarding transparency when a cut-out is requested and full coverage when an opaque image is requested. Let denote the CLIP-style prompt–image cosine similarity computed using embeddings from Qwen3-VL [bai2025qwen3vl]. Following the MLLM implicit feedback of UniWorld-v2 [li2025uniworld], let be the logit-based score from Qwen3.5-9B [qwen2026qwen35]. We prompt the frozen MLLM to rate the white-composited asset against its text condition on a scale from to , apply a softmax to the output logits of the six score tokens, and use their normalized expected value, .

For I2L, every decoded layer is paired by its ordered slot with the ground-truth RGBA layer on the same canvas. The first reward head is the mean absolute error over all four RGBA channels and all pixels. The second is a frozen AlexNet LPIPS network [zhang2018lpips] applied only to the three RGB channels, without an alpha mask:

The rewards are

(6)

The outer alpha-following term is a gate: a sample cannot compensate for incorrect opacity merely by receiving high embedding or MLLM scores. For I2L, the two explicit reconstruction distances are first combined within each layer and then averaged equally across target layers; the stable configuration uses equal and LPIPS weights. We do not multiply rewards across layers and use no reward on the recomposed image. Alpha errors therefore affect the RGBA head but never enter the perceptual network.

4 Evaluation

4.1 Setup

We evaluate I2L by per-layer fidelity and editability, and T2RGBA by text alignment, reference resemblance, and alpha quality.

Image-to-Layer.

We use fixed-seed Crello designs [yamaguchi2021canvasvae], pairing each rendered design with its source stack; no Crello data are used for training. Our final checkpoint runs at -equivalent image area, eight steps, and no classifier-free guidance. We request four layers to match Qwen-Image-Layered, which runs through its official script [yin2025qwenimagelayered] at the recommended pixels with its optional global-caption input.

LayerD [suzuki2025layerd] supplies stack alignment and per-layer RGB/alpha metrics; Stable-Layers [rowles2026stablelayers] supplies reference-free editability and the structural rubric. Definitions and resampling rules are in Appendix B; each is one standard deviation across samples.

Text-to-RGBA.

From an internally curated held-out set, we sample captioned layers at seed and generate one asset per caption. The evaluation captions, their matched source layers and the real-layer FID reference set do not overlap the training data. CLIP Score uses all outputs; FID [heusel2017fid] compares their white composites with real layers. Alpha MSE, SAD, and white-composite LPIPS use the locatable source layers at their native resolutions; availability is fixed before inference and shared by all methods. We compare LayerDiffuse [zhang2024layerdiffuse] and OmniAlpha [yu2025omnialpha].

4.2 Layer Decomposition Quality

媒体内容 · 前往原文查看

Table 1: Per-layer quality on 512 Crello designs. LayerD dynamic-time-warping alignment is applied before scoring; RGB L1 uses binary reference-alpha support.

Method RGB L1 Alpha Soft IoU

Qwen-Image-Layered ( layers)

UniWorld-I2L ( layers)

At matched layer count, Table 1 shows a relative reduction in per-layer RGB L1 () and a relative increase in Alpha Soft IoU (). Thus, after stack alignment, our layers more closely match both the RGB content and alpha support of their references.

媒体内容 · 前往原文查看

Table 2: Editability on 512 Crello designs. Stable-Layers reference-free metrics; Bad Layers counts blank and glazed layers per sample, and the percentage is per emitted slot.

Method Bad Layers Blank Glaze Feat. Dist.

Qwen-Image-Layered ( layers) ()

UniWorld-I2L ( layers) ()

The reference-free metrics in Table 2 are consistent with the per-layer fidelity results in Table 1. Blank outputs fall by (), and Feature Distribution rises from to , suggesting less redundant content. Glazed outputs also fall from to , reducing Bad Layers from to .

媒体内容 · 前往原文查看

Table 3: VLM-based evaluation of four-layer decomposition following the Stable-Layers protocol. Five dimensions are scored from to ; values are mean standard deviation across samples. The total is out of , with its normalized value in parentheses.

Method Semantic separation Alpha cleanliness Background inpainting Content distribution Content validity Total (normalized)

Qwen-Image-Layered ()

UniWorld-I2L ()

The VLM evaluation in Table 3 shows improvements in four dimensions and a decline in alpha cleanliness. Content distribution (), content validity (), and semantic separation () account for most of the -point total improvement; background inpainting adds . The latter is compatible with the effect expected from supervision using complete objects in real documents, although we do not isolate that effect here. Alpha cleanliness is the only lower dimension ( versus ) and is discussed in Section 5.

Qualitative comparison.

Figure 5 shows the same structural pattern on individual designs, while Figure 6 demonstrates that a returned layer can be decomposed again by a second call.

Figure 5: Qualitative comparison of per-layer decomposition with Qwen-Image-Layered on two Crello designs. For each design, the Qwen-Image-Layered row appears above ours; each row shows four predicted RGBA layers, through , ordered back to front. Checkerboard denotes transparency.

Figure 6: Qualitative analysis of instruction-controlled I2L decomposition. For each input poster, the first round (R1) follows the user prompt shown below the example to decompose the image into four semantic, position-aligned RGBA layers: background, subject, design, and text. The outlined subject layer becomes the input image for the second round (R2). R2 then follows its own user-specified prompt to decompose that selected layer into two finer-grained layers, again preserving position and alpha.

Figure 6: Qualitative analysis of instruction-controlled I2L decomposition (continued).

Figure 6: Qualitative analysis of instruction-controlled I2L decomposition (continued).

4.3 Text-to-RGBA Generation

媒体内容 · 前往原文查看

Table 4: Text-to-RGBA generation on 512 design-element prompts. FID and CLIP Score use all generations, with FID computed against real layers. Alpha MSE, SAD [levin2008matting], and LPIPS use the prompts with matched reference layers, evaluated at the reference resolution (Section 4.1, Appendix B). OmniAlpha uses its released reinforcement-learning LoRA at , steps, and true guidance .

Method FID CLIP Score Alpha MSE SAD LPIPS (white)

LayerDiffuse (SDXL)

OmniAlpha

UniWorld-T2RGBA

Figure 7: Text-to-RGBA against LayerDiffuse and OmniAlpha. The top panel contains seven single-resource prompts and the bottom panel seven multi-resource prompts. Each column uses the same prompt for all three models; the full English prompts are listed in Appendix D.

In Table 4, the improvement over LayerDiffuse is consistent: FID falls by , CLIP Score rises by , and the three paired measures improve by –. Our model has the highest mean CLIP Score ( versus , ), whereas OmniAlpha leads on FID ( versus ) and white-composite LPIPS ( versus ). Its Alpha MSE and SAD are also lower, but by small margins ( versus and versus , respectively).

These metrics answer different questions. Prompt–image evaluation ranges from embedding alignment to knowledge-informed semantic criteria [radford2021clip, niu2025wise]; in our protocol, CLIP Score serves as the alignment proxy, while FID and LPIPS compare outputs with reference assets at the distribution and instance levels, respectively. Within this protocol, OmniAlpha better matches the reference appearance distribution and achieves slightly lower Alpha MSE and SAD, while our model follows the text more closely on average. Figure 7 complements these aggregate metrics with side-by-side generations for selected single-resource and multi-resource design briefs.

5 Limitations

Alpha edges.

Table 3 reports lower alpha cleanliness for UniWorld-I2L than for Qwen-Image-Layered ( versus ). Improving fine alpha boundaries remains future work.

Dense typography, and Chinese text in particular.

Complex glyphs and longer passages produce missing strokes, wrong characters and unstable layouts. Typography-intensive design will require stronger text generation in the underlying model.

6 Conclusion

UniWorld-Design treats semantic RGBA layers as native units of generation and interaction rather than outputs of a post-processing pipeline. T2RGBA generates reusable standalone RGBA assets, while I2L decomposes finished images into ordered, complete semantic layers through top-level decomposition, recursive refinement and targeted extraction. These layers form persistent, language-addressable object states that external design agents can select, manipulate with other tools and recompose.

Our evaluations show that I2L improves per-layer fidelity and usability over Qwen-Image-Layered, while T2RGBA achieves the highest mean CLIP Score among the compared models. Alpha-boundary quality and dense typography remain the primary limitations observed in our evaluations.

Next, we will scale data, model capacity and compute to pretrain a unified model for native RGBA generation and understanding, extending its scope from RGBA asset generation and image decomposition to layer insertion and replacement, layout-constrained composition and iterative multi-turn editing.

References

Appendix A Contributions

Contributors: Zongjian Li, Chenxu Bai, Chen Chen, Haoxiang Sun, Shaodong Wang, Feize Wu, Shenghai Yuan, Bin Lin, Zheyuan Liu, Yuwei Niu.

Project Leads: Zhiyuan Yan, Li Yuan† †††Corresponding author.

Appendix B Evaluation Details

Per-layer metrics (LayerD).

Predicted and reference layer stacks are treated as ordered sequences and aligned by the LayerD dynamic-time-warping edit protocol, which may merge adjacent layers before scoring to accommodate granularity differences. Under that alignment, each matched pair is compared at the smaller of the two resolutions (downsampling only). For matched pairs with predicted alpha , reference alpha , binary reference support , predicted RGB and reference RGB , all with RGB and alpha values in ,

(7)

(8)

both with values in . RGB L1 is weighted by the binary reference support, so a layer is judged where the reference alpha is nonzero, and Alpha Soft IoU uses continuous alpha rather than a binarized mask, which makes it sensitive to semi-transparent regions.

Editability metrics.

Following Stable-Layers, we compute these metrics using only the predicted layers. A layer is blank when its mean alpha is below . A layer is glazed when its mean alpha lies in and its RGB standard deviation exceeds — a diffuse semi-transparent layer containing residual color, which limits its independent editability. Bad Layers is the sum of the two counts. Feature Distribution extracts a DINOv2-base feature per layer (the spatial mean of the last hidden state) and reports , so a higher value means the layers of one output are less alike.

VLM judging.

GPT-5.6-terra receives the input design, the alpha composite of the predicted stack and a labeled contact sheet of the predicted layers, and returns five integers in : semantic separation (does each foreground layer contain a single complete element rather than a fragment), alpha cleanliness (are transparent boundaries free of color residue, halos and haze), background inpainting (is the region behind the foreground plausibly restored in ), content distribution (is content spread meaningfully across layers rather than concentrated in one), and content validity (does the recomposition match the input, with no missing content, blank layers or noise layers).

T2RGBA metrics.

FID uses Inception-v3 features at after compositing on white and compares generated samples with a reference distribution constructed from real layers. CLIP Score is the cosine similarity between the generated asset and its prompt under CLIP ViT-B/32, over the same generations. Alpha MSE, SAD and white-composite LPIPS are computed at the native resolution of the corresponding reference layer, using the prompts with successfully matched source layers (Section 4.1); all three methods are scored on that same subset. Alpha MSE is the mean squared error between the predicted and reference alpha channels on , and SAD [levin2008matting] is the classical matting statistic over the reference layer’s own pixels. White-composite LPIPS uses an AlexNet backbone.

Appendix C Sequence Construction

Token packing.

Tokens are packed in layer-major order — every token of layer , then every token of layer , and so on — rather than interleaving layers at each spatial position. Combined with a shared crop and resize across the layers of a sample, this makes a given refer to the same canvas location in every layer. Text is packed as a global instruction followed, for each target layer , by Layer : <short> --- <detailed>; the explicit index is what binds a description to a layer slot and lets the layer count be set at inference.

Appendix D Prompts for the Qualitative T2RGBA Comparison

Figure 7 uses seven single-resource prompts in its top panel and seven multi-resource prompts in its bottom panel. The prompts below follow the figure’s left-to-right column order.

Single-resource prompts (top panel).

Prompt. one peach colored cat reading a tiny sky blue book only, peach, sky blue, vanilla palette, refined editorial gouache texture, delicate colored-pencil detail, soft controlled highlights, crisp clean silhouette, subject occupies 70 percent of canvas, centered, transparent background

Prompt. one complete round red wax seal stamp with embossed postal bird emblem, full seal visible only, postal red, indigo, aged paper palette, refined editorial gouache texture, delicate colored-pencil detail, soft controlled highlights, crisp clean silhouette, subject occupies 70 percent of canvas, centered, transparent background

Prompt. one kraft canvas market tote filled with colorful vegetables only, tomato red, leafy green, sunflower yellow palette, refined editorial gouache texture, delicate colored-pencil detail, soft controlled highlights, crisp clean silhouette, subject occupies 70 percent of canvas, centered, transparent background

Prompt. one kawaii turquoise whale wearing a tiny yellow sailor hat only, turquoise, lemon, bubblegum pink palette, refined editorial gouache texture, delicate colored-pencil detail, soft controlled highlights, crisp clean silhouette, subject occupies 70 percent of canvas, centered, transparent background

Prompt. one rose pink flower bouquet wrapped in kraft paper only, rose pink, leaf green, kraft brown palette, refined editorial gouache texture, delicate colored-pencil detail, soft controlled highlights, crisp clean silhouette, subject occupies 70 percent of canvas, centered, transparent background

Prompt. one complete lime green hand citrus juicer with ribbed bowl and short handle, full juicer visible only, lime green, orange, clean white palette, refined editorial gouache texture, delicate colored-pencil detail, soft controlled highlights, crisp clean silhouette, subject occupies 70 percent of canvas, centered, transparent background

Prompt. one two tier birthday cake with pink frosting and blue candles only, confetti pink, bright blue, golden yellow palette, refined editorial gouache texture, delicate colored-pencil detail, soft controlled highlights, crisp clean silhouette, subject occupies 70 percent of canvas, centered, transparent background

Multi-resource prompts (bottom panel).

Prompt. Create a polished cozy breakfast sticker sheet as a cohesive hand-journal and scrapbook design asset. Include croissant, jam jar, coffee cup, gingham napkin, tiny sparkles. Use a butter yellow, strawberry red, oat beige palette, charming hand-drawn gouache and colored-pencil texture, delicate paper grain, clean white sticker borders, subtle layered paper shadows, balanced spacing, and varied scale. Arrange the pieces as one compact editorial sticker cluster, isolated on a truly transparent background, no mockup, no watermark, no extra text, crisp cutout edges, premium commercial design quality.

Prompt. Create a polished spring gardening scrapbook kit as a cohesive hand-journal and scrapbook design asset. Include watering can, seed packet, tulips, garden gloves, paper labels. Use a sage green, peach, cream palette, charming hand-drawn gouache and colored-pencil texture, delicate paper grain, clean white sticker borders, subtle layered paper shadows, balanced spacing, and varied scale. Arrange the pieces as one compact editorial sticker cluster, isolated on a truly transparent background, no mockup, no watermark, no extra text, crisp cutout edges, premium commercial design quality.

Prompt. Create a polished rainy day journaling cluster as a cohesive hand-journal and scrapbook design asset. Include clear umbrella, rain boots, cloud, tea mug, dotted washi tape. Use a dusty blue, lavender, warm gray palette, charming hand-drawn gouache and colored-pencil texture, delicate paper grain, clean white sticker borders, subtle layered paper shadows, balanced spacing, and varied scale. Arrange the pieces as one compact editorial sticker cluster, isolated on a truly transparent background, no mockup, no watermark, no extra text, crisp cutout edges, premium commercial design quality.

Prompt. Create a polished seaside vacation ephemera set as a cohesive hand-journal and scrapbook design asset. Include shells, striped deck chair, postcard, sunglasses, postage stamps. Use a aqua, coral, sand palette, charming hand-drawn gouache and colored-pencil texture, delicate paper grain, clean white sticker borders, subtle layered paper shadows, balanced spacing, and varied scale. Arrange the pieces as one compact editorial sticker cluster, isolated on a truly transparent background, no mockup, no watermark, no extra text, crisp cutout edges, premium commercial design quality.

Prompt. Create a polished city weekend travel ephemera as a cohesive hand-journal and scrapbook design asset. Include subway ticket, camera, skyline, map pin, street sign. Use a cobalt, taxi yellow, charcoal palette, charming hand-drawn gouache and colored-pencil texture, delicate paper grain, clean white sticker borders, subtle layered paper shadows, balanced spacing, and varied scale. Arrange the pieces as one compact editorial sticker cluster, isolated on a truly transparent background, no mockup, no watermark, no extra text, crisp cutout edges, premium commercial design quality.

Prompt. Design a high-converting summer promotion decorative type for a modern e-commerce campaign. The exact main lettering must read "SUMMER SAVINGS". Use a aqua, coral, sunny yellow palette with sun, wave, tropical leaves, bubbly dimensional letters; create bold custom display typography, clear hierarchy, dimensional highlights, clean shadows, energetic but controlled ornament, and excellent readability at thumbnail size. Isolate one complete button or badge asset on a truly transparent background, no product photo, no mockup, no watermark, no extra words, crisp alpha edges, premium agency-level commercial finish.

Prompt. Design a high-converting app download promotional badge for a modern e-commerce campaign. The exact main lettering must read "DOWNLOAD APP". Use a gradient blue, purple, white palette with smartphone, downward arrow, pixel sparkles; create bold custom display typography, clear hierarchy, dimensional highlights, clean shadows, energetic but controlled ornament, and excellent readability at thumbnail size. Isolate one complete button or badge asset on a truly transparent background, no product photo, no mockup, no watermark, no extra words, crisp alpha edges, premium agency-level commercial finish.