OminPSD: Layered PSD Generation with Diffusion Transformer

Cheng Liu, Yiren Song, Haofan Wang, Mike Zheng ShouCVPR 2026 Findings 2026 · OpenReview

TL;DR

OmniPSD is a unified diffusion-transformer framework that bridges the gap between flat image generation and professional graphic design by generating and decomposing fully editable, multi-layer Photoshop (PSD) files. By integrating a custom RGBA-VAE with the Flux ecosystem, it handles both text-to-PSD generation and image-to-PSD decomposition while natively preserving transparent alpha channels and structural consistency.

What Problem Does This Paper Address?

While modern generative models excel at producing stunning, high-fidelity flat images, they fall short when integrated into professional design workflows. Designers rely heavily on layered formats like Photoshop (PSD) files, which separate elements (text, foreground objects, backgrounds) into distinct, transparent layers for flexible editing.

Existing attempts to generate layered content usually suffer from two major flaws: they either perform post-hoc decomposition on a flat RGB image (leading to error accumulation and weak consistency), or they lack true alpha-channel (transparency) awareness, failing to handle complex design elements like semi-transparent text, soft shadows, and edge blending. OmniPSD addresses this by treating layered generation and decomposition as a unified, bidirectional task within a single architecture.

OmniPSD pipeline overview for multi-layer synthesis and reconstruction

Methods

OmniPSD is built upon the Flux diffusion transformer ecosystem and introduces three core components to enable transparency-aware, layered manipulation:

  1. RGBA-VAE: The authors retrained a unified variational autoencoder (extending AlphaVAE) on real-world design samples. This provides a shared latent space that accurately encodes and decodes both color and alpha channels, preserving soft edges and complex blending.
  2. Text-to-PSD via Spatial In-Context Learning: Instead of modifying the model architecture to handle multiple layers, OmniPSD arranges target layers (full poster, foreground, midground, background) into a 2×2 grid canvas. Driven by hierarchical JSON prompts, the model leverages its native spatial self-attention mechanism to implicitly learn cross-layer relationships, occlusion ordering, and color harmony in a single forward pass.
  3. Image-to-PSD via Iterative Decomposition: For reverse-engineering flat images, OmniPSD uses Flux-Kontext with task-specific LoRA adapters. It operates an iterative loop: it extracts a foreground/text layer, erases it to restore the occluded background underneath, and repeats. Additionally, an integrated OCR and font-recovery pipeline converts rasterized text back into resolution-independent, editable vector text layers.
Detailed overview of the OmniPSD architecture for training and inference

Datasets

To train and evaluate the framework, the authors constructed the Layered Poster Dataset, containing over 200,000 professional PSD files collected from online design repositories.

  • The dataset features manual groupings, font layers, shape groups, and effect overlays.
  • Automated parsing and strict post-filtering were applied to ensure valid RGBA structures.
  • Each sample is accompanied by structured JSON metadata detailing bounding boxes, visibility, stacking order, and layer-specific textual descriptions.
The structure of the Layered Poster Dataset for both sub-tasks

Key Insights and Ideas

  • The 2×2 Grid Trick: One of the most elegant insights is using a spatial grid as an in-context canvas for multi-layer generation. By stitching layers into a single image, the transformer's standard spatial attention naturally acts as a cross-layer reasoning mechanism, bypassing the need for complex, custom multi-branch architectures.
  • Iterative "Extract & Erase" Loop: Decomposing an image all at once is highly error-prone. Breaking it down into an alternating sequence of foreground extraction and background inpainting mirrors the manual workflow of a human designer and yields significantly cleaner alpha boundaries.

Key Results

OmniPSD was evaluated against strong baselines like LayerDiffuse and GPT-Image-1, demonstrating substantial improvements across both quantitative metrics and qualitative visual fidelity.

  • Text-to-PSD Generation: OmniPSD achieved an FID score of 30.43, vastly outperforming LayerDiffuse SDXL (89.35) and GPT-Image-1 (53.21), while also leading in semantic alignment (CLIP Score of 37.64).
  • Image-to-PSD Decomposition: In reconstruction fidelity, OmniPSD achieved a remarkably low Mean Squared Error (MSE of 1.14e-3) and a high PSNR of 24.0, proving its ability to accurately reconstruct flat images back into their original constituent layers.
Qualitative results showing clean layer separation and background restoration

Limitations and Future Work

The limitations and specific directions for future work are not explicitly discussed in the provided text. However, based on the architecture, potential areas of exploration include scaling the framework to support an arbitrary number of layers beyond the fixed 2×2 grid structure and extending the framework to handle more complex vector graphic shapes natively.

Summary updated

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 imageto- 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 RGBAVAE 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. Project page: https://showlab.github.io/OmniPSD/.
BibTeX
@inproceedings{liu2026ominpsd,
  author = {Cheng Liu and Yiren Song and Haofan Wang and Mike Zheng Shou},
  title = {{OminPSD: Layered PSD Generation with Diffusion Transformer}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026},
  url = {https://openreview.net/forum?id=3kJA3hFWvD},
}

Loading related content...

Figure