OmniPSD: Layered PSD Generation with Diffusion Transformer (2512.09247v1)

Abstract: 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.

• The paper introduces OmniPSD, a unified architecture that integrates transformer-based diffusion models to generate and decompose layered PSD files with accurate transparency handling.
• The methodology employs a shared RGBA-VAE and hierarchical prompts to ensure coherent, editable layer synthesis and reconstruction.
• Experimental results demonstrate significant improvements in metrics like MSE, PSNR, and SSIM, surpassing existing flat image generation methods.

OmniPSD: Layered PSD Generation with Diffusion Transformer

Framework Overview and Objectives

OmniPSD introduces a unified architecture for bidirectional layered design, capable of both text-to-PSD multi-layer synthesis and image-to-PSD reconstruction (Figure 1). The framework utilizes a shared RGBA-VAE to encode transparent-layered images in a latent space, enabling accurate modeling of layer structure and alpha-channel transparency. The core system integrates two complementary transformer-based diffusion models, Flux-dev for generative synthesis and Flux-Kontext for in-context decomposition, all within the Flux ecosystem. The solution is targeted at bridging the gap between existing generative visual models—which typically output only flat raster images—and professional design workflows that require transparent, hierarchically compositional PSD assets.

Figure 1: OmniPSD architecture supports both text-driven synthesis and in-context decomposition of layered PSD files with full RGBA support.

The RGBA-VAE module provides high-fidelity latent encoding and decoding of alpha channels, overcoming limitations of previous VAEs that were pre-trained only on natural transparency data. This is critical for representing semi-transparent textual overlays, compositional effects, and nuanced blending in design-centric contexts (Figure 2).

Figure 2: RGBA-VAE achieves superior alpha-channel reconstruction compared to competing VAE approaches.

Dataset and Training Paradigm

OmniPSD’s training is anchored on a substantial dataset of 200,000 professionally-authored PSD files, parsed to extract semantic groupings—text, foreground, background—with explicit RGBA structure. For text-to-PSD, the dataset is formatted into a $2 \times 2$ grid combining the full poster and its decomposed layers, enabling spatial in-context learning. For image-to-PSD, the triplet strategy emulates iterative decomposition: input image, extracted foreground, and post-erasure background (Figure 3).

Figure 3: Layered dataset format for supervised learning of both text-to-PSD synthesis and iterative image-to-PSD decomposition.

Image-to-PSD Decomposition via Iterative In-Context Editing

The image-to-PSD reconstruction pipeline is formulated as a sequence of LoRA-adapted transformer modules applied iteratively. Text and foreground elements are extracted and erased against the remaining background, with each step producing RGBA PNG assets. This deterministic decomposition approach, modeled with Flow Matching loss in latent space, yields pixel-precise layer separation and composability. Editable text layers are recovered using transformer-based OCR with semantic font retrieval, producing editable vector layers rather than rasterized text—essential for workflow interoperability.

The decomposition process is especially robust, outperforming commercial solutions and prior non-alpha-aware segmentation methods in all key metrics, including layer-wise FID, MSE, PSNR, SSIM, CLIP-I, and GPT-4-based spatial consistency scores. Table results demonstrate an MSE reduction by over an order of magnitude and state-of-the-art perceptual similarity for full reconstruction.

Text-to-PSD Multi-Layer Synthesis via Spatial In-Context Reasoning

For forward generation, OmniPSD employs hierarchical prompts, assigning independent descriptions for each design layer. The $2 \times 2$ grid canvas allows the DiT backbone’s spatial attention to implicitly model compositional and occlusion relationships, with no need for additional cross-layer modules. This yields a single-pass generation of all PSD layers, maintaining logical stacking and semantic consistency.

Qualitative and quantitative evaluations (Figure 4 and Figure 5) show that OmniPSD matches or surpasses baseline models—including LayerDiffuse and GPT-Image-1—in overall FID, semantic CLIP alignment, and spatial GPT-4 scores. Notably, OmniPSD uniquely delivers logically-coherent, transparency-preserving RGBA layers that are directly editable in professional design tools.

Figure 4: OmniPSD generates and reconstructs clear, editable PSD layers in both synthesis and decomposition tasks.

Figure 5: OmniPSD results demonstrate superior layer fidelity and logical separation compared to all competing baselines.

Ablation and Component Analysis

An ablation study reveals decisive gains attributable to the retrained RGBA-VAE and structured layer-specific prompts. Relative to previous VAE models trained on generic data, the RGBA-VAE provides marked improvements in reconstruction metrics, including MSE, PSNR, SSIM, and perceptual LPIPS. Furthermore, the use of structured hierarchical prompts in text-to-PSD synthesis yields consistent improvements in both quantitative and qualitative evaluations.

Expanded Qualitative Results and Practical UI Demonstrations

Additional visualizations exhibit the clarity and compositional realism of decomposed layers and synthesized multi-layer posters (Figure 6, Figure 7). In practice, OmniPSD enables designers to upload raster images and instantly derive editable PSD stacks—text, foreground, background—which can be modified, rearranged, or extended using an intuitive user interface (Figure 8).

Figure 6: Further examples of OmniPSD’s accurate layered decomposition from flattened poster images.

Figure 7: Examples of diverse text-to-PSD generation with logical layer stacking and clean alpha boundaries.

Figure 8: OmniPSD UI enables granular editing—adding, removing, and modifying text and graphics—in decomposed layered assets.

Implications and Future Directions

OmniPSD demonstrates that unified architectures can achieve bidirectional translation between raster images, text prompts, and editable PSD assets, all under an RGBA- and transparency-aware regime. This not only sets a new technical standard for design-relevant image generation and decomposition, but also enables new workflows for automated asset creation, rapid decoupled editing, and layer-wise reasoning in both generative and reconstructive pipelines.

From a theoretical standpoint, OmniPSD’s approach validates the utility of spatial in-context learning and hierarchical reasoning in transformer-based diffusion models for compositional tasks. The shared latent space and alpha-aware representation align closely with requirements for professional design systems.

Leading questions for future research include extending the model to vector graphics decomposition, animated multi-layer video synthesis, and more granular semantic control in both the prompt and layer domains. The paradigm introduced here can be expected to proliferate toward more general multi-modal, highly editable design generation settings, and inform further advances in compositional generative models.

Conclusion

OmniPSD establishes a robust, unified framework for layered, alpha-preserving PSD generation and decomposition, built on transformer-based diffusion and supported by high-fidelity RGBA-VAE representation learning. Through extensive evaluation, OmniPSD sets new benchmarks for editable, structurally consistent design asset synthesis and analysis, opening both practical and theoretical avenues in cross-modal, in-context aware generative modeling.

Reference: "OmniPSD: Layered PSD Generation with Diffusion Transformer" (2512.09247).