PaperlayerView exact revision on arXiv

Source-linked AI summary

OmniPSD: Layered PSD Generation with Diffusion Transformer

Cheng Liu, Yiren Song, Haofan Wang, Mike Zheng Shou

arXiv:2512.09247v1cs.CV

TL;DR

Layered PSD generation and reconstruction remain challenging because most generative models output flat raster images without layer structure or transparency. OmniPSD combines a shared RGBA-VAE with Flux-based in-context generation and decomposition workflows. The authors report high-fidelity generation and reconstruction with accurate alpha representation, presenting the framework as supporting both creative synthesis and structural reconstruction.

  • Problem

    Most generative models output flat raster images, lacking the layer-wise structure and transparency needed for professional design workflows.

    Source: §1 · p. 1

  • Method

    OmniPSD combines a shared RGBA-VAE with Flux-based workflows for spatially arranged text-to-PSD generation and iterative image-to-PSD decomposition.

    Source: §1 · pp. 1–2

  • Results

    The authors report high-fidelity generation and reconstruction with accurate alpha-channel representation, supported by extensive experiments.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper presents transparency-aware in-context learning as a framework for both creative synthesis and structural reconstruction.

    Source: §1 · p. 2

Abstract

from arXiv · show

Recent advances in diffusion models have greatly improved image generation and editing, yet generating or reconstructing layered PSD files with transparent alpha channels remains highly challenging. We propose OmniPSD, a unified diffusion framework built upon the Flux ecosystem that enables both text-to-PSD generation and image-to-PSD decomposition through in-context learning. For text-to-PSD generation, OmniPSD arranges multiple target layers spatially into a single canvas and learns their compositional relationships through spatial attention, producing semantically coherent and hierarchically structured layers. For image-to-PSD decomposition, it performs iterative in-context editing, progressively extracting and erasing textual and foreground components to reconstruct editable PSD layers from a single flattened image. An RGBA-VAE is employed as an auxiliary representation module to preserve transparency without affecting structure learning. Extensive experiments on our new RGBA-layered dataset demonstrate that OmniPSD achieves high-fidelity generation, structural consistency, and transparency awareness, offering a new paradigm for layered design generation and decomposition with diffusion transformers.

1. Introduction

OmniPSD unifies text-to-PSD generation and image-to-PSD decomposition in one diffusion framework, producing editable, multi-layer graphics with transparency. It combines a shared RGBA-VAE with Flux models and task-specific in-context workflows.

1. Introduction

  • OmniPSD supports text-to-PSD generation and image-to-PSD decomposition within one diffusion-based architecture, bridging creative generation and analytical reconstruction.The framework transforms textual or visual inputs into editable, multi-layer PSD graphics.Source: §1 · pp. 1–2
  • A shared, pre-trained RGBA-VAE preserves alpha information across Flux-dev generation and Flux-Kontext in-context editing.The RGBA-VAE encodes and decodes transparent images in a latent space that retains alpha-channel information.Source: §1 · p. 1
  • Text-to-PSD places background, foreground, text, and effects in a 2 × 2 grid for simultaneous generation, while image-to-PSD iteratively extracts text, restores the background, and refines foreground layers.The generated and decomposed layers are decoded or output in RGBA format for transparent boundaries and compositional relationships.Source: §1 · p. 2
  • The paper presents the unified approach as showing that creative synthesis and structural reconstruction can both be achieved through transparency-aware in-context learning.Source: §1 · p. 2
  • The authors report high-fidelity generation and reconstruction with accurate alpha representation, and introduce a large-scale dataset with detailed RGBA layer annotations and a benchmark.They report that extensive experiments demonstrate the approach's effectiveness and superiority.Source: §1 · p. 2

2. Related Works

Diffusion models have advanced from U-Net-based denoisers toward Transformer backbones and support a broad range of visual tasks. Layered-image research increasingly targets direct multi-layer generation and transparency-aware representations, while many graphic-design systems produce flattened or coarse-vector outputs rather than PSD-style RGBA layers.

2.1. Diffusion Models

  • Diffusion models span text-to-image generation, editing, and video synthesis; Transformer denoisers have become a common backbone, while flow-matching and ODE formulations learn continuous deterministic flows.Earlier work predominantly used U-Net denoisers in pixel or latent space; Transformer models use global attention and scalability to improve visual fidelity and prompt alignment.Source: §2.1 · p. 2

2.2. Layer-wise Image Generation

  • Early image decomposition focused on depth layers, alpha mattes, or semantic regions, while post-hoc segmentation and generate-then-decompose pipelines can accumulate errors and offer limited control over layout and inter-layer relationships.Source: §2.2 · p. 2
  • LayerDiff, LayerDiffuse, and LayerFusion jointly synthesize RGBA layers and model occlusion, while ART and Layer-Tracer target variable layouts and object-level control [11] [25] [75] [48] [57].MuLAn provides RGBA annotations and occlusion labels; PSDiffusion harmonizes foreground and background layout and appearance [60] [24].Source: §2.2 · pp. 2–3
  • Graphic-design systems use hierarchical pipelines, structured layouts, dual raster-vector generation, or markup and templates, but typically output flattened renders or coarse vector structures rather than explicit PSD-style RGBA layers [27] [28] [10] [31] [66] [71].The cited passage describes PSD-style layers with explicit alpha channels as directly editable and composable in professional design tools.Source: §2.2 · p. 3

2.3. RGBA Image Generation

  • Traditional transparency workflows add separate matting, segmentation, or alpha-estimation steps to opaque RGB generation, while newer work models transparency in diffusion and RGBA-aware generation [7] [23] [53] [68] [75] [15] [49].Related systems also combine RGBA autoencoders and diffusion transformers to model transparency and inter-layer effects, often using matting-centric datasets [64] [9] [11] [24] [48] [65] [70] [74] [60].Source: §2.3 · pp. 3–4

3. Method

OmniPSD combines Flux generation and editing models with a shared alpha-aware latent space to generate layered designs from text and decompose flattened images. Its method includes iterative image editing, spatial in-context layer learning, and task-specific dataset construction.

3.1. Overall Architecture

  • OmniPSD unifies Flux-Dev generation and Flux-Kontext editing with a shared RGBA-VAE for bidirectional conversion between design images and editable PSD layers.The two branches use iterative editing for decomposition and spatial attention for generation.Source: §3.1 · p. 4

3.2. RGBA-VAE

  • Retraining AlphaVAE on real design samples addresses degradation on semitransparent text, shadows, and soft blending, providing a shared latent space for alpha-preserving generation and decomposition.The objective balances pixel fidelity, patch-level consistency, perceptual alignment, and latent regularization.Source: §3.2 · p. 4

3.3. Image-to-PSD Reconstruction

  • Image-to-PSD progressively extracts text and foreground objects, erases foregrounds to restore occluded backgrounds, and outputs RGBA layers for PSD reconstruction.The pipeline uses task-specific LoRA adapters; its flow-matching formulation maps encoded input images toward target layer representations.Source: §3.3 · p. 5
  • A unified OCR, font-recovery, and rendering pipeline turns rasterized text into editable vector-text layers that preserve typography and layout.Recognized text is matched to a typeface from a curated font bank before being re-rendered.Source: §3.3 · p. 5

3.4. Text-to-PSD Generation

  • Text-to-PSD places the full poster and three decomposed layers in a 2×2 grid, using spatial attention and hierarchical prompts to learn cross-layer relationships without added losses or explicit cross-layer modules.The model generates all PSD layers jointly in one pass.Source: §3.4 · pp. 5–6

3.5. Dataset Construction

  • The Layered Poster Dataset contains over 200,000 real PSD files with RGBA layer annotations and editable metadata; text layers are omitted from Text-to-PSD samples, while Image-to-PSD uses input, foreground, and background triplets.Text-to-PSD samples use a four-panel grid, and Image-to-PSD triplets mirror iterative extraction and background restoration.Source: §3.5 · p. 6

4. Experiments

Experiments assess text-to-PSD generation and image-to-PSD reconstruction using real-PSD test sets, quantitative metrics, and qualitative comparisons. OmniPSD is reported to produce coherent, editable layers, while ablations show that poster-suited VAEs and structured prompts matter.

4.1. Experiment Details.

  • Evaluation uses 500 layer-aware prompts for text-to-PSD and 500 flattened real PSD images for reconstruction, with task-specific metrics and GPT-4 layout-coherence scoring.Training uses 1024×1024 resolution; text-to-PSD uses a 2×2 grid, while image-to-PSD trains separate foreground extraction and erasure modules.Source: §4.1 · pp. 6–7

4.2. Comparison and Evaluation

  • OmniPSD matches leading diffusion and vision-language models in visual quality while uniquely supporting transparent multi-layer PSDs, with stronger fidelity and layer coherence than existing layered synthesis baselines.Qualitative comparisons report coherent posters and accurate text extraction, foreground extraction and removal, and background reconstruction.Source: Figure 5 · p. 8, §4.2 · p. 8

4.3. Ablation Study.

  • VAEs trained primarily on natural images reconstruct design posters inconsistently, with artifacts and blurred text; naive prompts significantly degrade text-to-PSD generation.The RGBA-VAE comparison is reported in Tables 4 and 6, and the prompt ablation in Table 5.Source: §4.3 · pp. 8–9, Table 4 · p. 9, Table 5 · p. 9

5. Conclusion

OmniPSD uses a Diffusion Transformer and iterative in-context editing to decompose raster posters into structured RGBA layers. The conclusion reports strong structural fidelity, transparency modeling, and semantic consistency.

5. Conclusion

  • OmniPSD combines an RGBA-VAE with task-specific Kontext-LoRA modules to decompose poster-style raster images into layers and reconstruct text, objects, and backgrounds.The paper also presents a large-scale, professionally annotated layered dataset for training and evaluation.Source: §5 · p. 9

A. GPT-4 Evaluation

The supplementary evaluation presents candidate layered posters together to GPT-4 under a shared input context. Scores use a five-level visual-quality rubric and are averaged and normalized for the reported metric.

A. GPT-4 Evaluation

  • GPT-4 evaluates all candidate methods for the same input together in one query, assigning each method an independent score for side-by-side comparison.The criteria include layer consistency, plausible occlusion and depth, and readability and layout of the composed poster.Source: A. GPT-4 Evaluation · p. 13
  • The rubric ranges from 1 for very poor results to 5 for very good results, with structure, layer relationships, and overall visual coherence guiding the ratings.The examples specify integer scores from 1 to 5.Source: A. GPT-4 Evaluation · p. 13
  • The reported GPT-4 metric averages integer ratings across test cases and linearly normalizes the result to [0, 1].Each method receives one score per query.Source: A. GPT-4 Evaluation · p. 13

B. User Study

The user study compares OmniPSD with LayerDiffuse-SDXL and GPT-Image-1 in both generation and decomposition. Participants rate layer quality and overall preference, and the study reports the highest mean scores for OmniPSD on the listed criteria.

B. User Study

  • In text-to-PSD evaluation, OmniPSD receives the highest mean scores for layering reasonableness and overall preference among the three methods.Participants rated 50 prompts on a five-point Likert scale.Source: B. User Study · p. 13, Table 6 · p. 13
  • In image-to-PSD evaluation, OmniPSD receives the highest mean scores for reconstruction consistency, layering reasonableness, and overall preference.Participants assessed 50 poster images using a five-point Likert scale.Source: B. User Study · p. 13, Table 7 · p. 14
  • Across both settings, designers praised clear layer separation and realistic transparency as supporting direct reuse in professional editing workflows.The conclusion characterizes the results as showing practical value for design generation and reconstruction.Source: B. User Study · pp. 13–14

C. More Results

OmniPSD's additional results cover both image-to-PSD reconstruction and text-to-PSD synthesis, with reported visual quality and layer properties in each setting.

C. More Results

  • Image-to-PSD reconstructions closely match original designs while preserving clean layer boundaries convenient for downstream editing.Source: C. More Results · p. 14
  • Text-to-PSD synthesis produces layered posters with coherent foreground elements, legible text, and visually consistent backgrounds.Source: C. More Results · p. 14
  • The additional qualitative examples span diverse layouts and contents, illustrating layer clarity, transparency and occlusion realism, and overall poster quality.Source: C. More Results · p. 14

D. User Interface

The interface starts from an uploaded poster and supports editing its inferred text, foreground objects, and background as separate components.

D. User Interface

  • From a user-uploaded image, OmniPSD infers editable layers separating text, foreground objects, and background regions.Source: D. User Interface · p. 14
  • Point-and-click operations let users modify text, remove or replace the background, and delete or adjust graphical elements.Source: D. User Interface · p. 14, Figure 9 · p. 17
  • These edits preserve the poster's overall layout and visual coherence while supporting practical poster editing and creation.Source: D. User Interface · p. 14

Related papers

Loading 2512.09247v1…