OmniPSD: AI Breakthrough Turns Flat Images into Editable Layers

In the realm of digital design, editing raster images has long been a frustrating ordeal. Once flattened into formats like JPEG or PNG, the distinct elements—text, foreground objects, and backgrounds—become inseparable, forcing designers into laborious manual rework. Now, a team from the National University of Singapore (NUS) Show Lab and Lovart AI has introduced OmniPSD, a groundbreaking AI model that shatters this limitation. By leveraging a unified diffusion-transformer architecture, OmniPSD can decompose a single flattened poster image into a layered Photoshop document (PSD) with transparent alpha channels or generate an editable layered design directly from text. This dual capability promises to automate and streamline design workflows, offering unprecedented flexibility for professionals and developers alike.

From Pixels to Layers: The Decomposition Revolution

At its core, OmniPSD excels at image-to-PSD decomposition. Given a flattened image, the model employs a diffusion-based flow-matching technique to iteratively separate elements into clean, editable RGBA layers. Unlike traditional methods relying on manual selection or brittle computer vision algorithms, OmniPSD's approach is rooted in generative AI, enabling it to understand semantic context and structural relationships within designs. The result is a PSD file where text, foreground elements, and backgrounds are isolated with transparency intact, ready for immediate editing in tools like Adobe Photoshop.

"OmniPSD delivers sharper reconstructions and cleaner separated text, foreground, and background layers, while better preserving the original layout and colors." — OmniPSD Research Team

This decomposition isn't merely about splitting images; it's about preserving design integrity. The model maintains color fidelity and spatial coherence, ensuring that reconstructed layers align perfectly—a critical requirement for professional workflows. For developers, this could simplify integration into design automation pipelines, enabling applications that transform user uploads or legacy assets into modifiable components.

From Text to Design: Synthesizing Editable Graphics

Equally transformative is OmniPSD's text-to-PSD synthesis capability. By feeding a text prompt into the model, users can generate entirely new layered PSDs. For instance, a prompt like "a minimalist tech poster with bold typography and geometric backgrounds" could produce a PSD with distinct layers for text, foreground graphics, and background elements. This is achieved through a 2×2 in-context RGBA grid and hierarchical captions within the diffusion transformer, which jointly generate layers while ensuring semantic coherence.

The synthesis process leverages the same unified architecture as decomposition, allowing the model to transfer knowledge between tasks. This shared foundation—particularly the RGBA-VAE (Variational Autoencoder)—ensures consistent handling of transparency and color across both modes. The result is more coherent layering and better-aligned editable text, addressing a common pain point in AI-generated design assets.

The Technical Backbone: Unified Diffusion-Transformer

OmniPSD's innovation lies in its unified Diffusion-Transformer architecture, which combines two powerful paradigms:

1. Diffusion-Based Flow-Matching: For decomposition, the model iteratively refines a noisy representation of the input image, gradually isolating layers. This approach ensures stability and high-fidelity reconstruction.
2. Transformer with Hierarchical Captions: For synthesis, the transformer processes text prompts to generate a hierarchical understanding of the desired design, translating it into structured layers.

The shared RGBA-VAE is pivotal, as it compresses and reconstructs images in the RGBA color space (including alpha channels), maintaining transparency throughout the pipeline. This unified design eliminates the need for task-specific models, reducing complexity while improving performance. Benchmarks show OmniPSD produces sharper results than prior methods, with cleaner text separation and superior layout preservation.

Implications for Design and Development

For designers, OmniPSD automates the tedious process of "unflattening" images, saving hours per project and enabling rapid prototyping. For developers, it opens doors to AI-driven design tools that can generate or parse layered assets programmatically. Imagine a CMS that automatically converts user-uploaded banners into editable PSDs, or a creative suite that generates marketing materials from text prompts.

Beyond immediate applications, OmniPSD advances research in structured generative AI. By controlling output as layered compositions rather than flat images, it demonstrates how diffusion models can handle complex, multi-element outputs—a challenge in fields like game asset creation or UI design. The open-source demo (available at Lovart AI) invites experimentation, potentially fostering community-driven improvements.

The Future of AI-Driven Design

OmniPSD signals a shift toward more intelligent, context-aware creative tools. As AI models evolve from generating static images to producing structured, editable assets, the line between raster and vector design blurs. This could democratize professional design workflows, allowing non-experts to create layered graphics while empowering professionals to iterate faster.

For the tech industry, OmniPSD underscores the potential of diffusion-transformer hybrids in solving real-world problems. By unifying decomposition and synthesis under one architecture, the research offers a blueprint for future models that generate complex, organized outputs—whether for design, 3D modeling, or beyond. As digital content demands grow, tools like OmniPSD will be pivotal in bridging the gap between human creativity and machine efficiency.