PSDesigner generates fully layered, editable Photoshop files from text

Anyone who has tried generating posters with AI knows the exact moment of frustration: the visual is almost perfect, but there is a typo in the main header, or the logo needs to move two inches to the left.

With a flat, one-layer image, you cannot just nudge the layer. Your only real choice is rolling the dice again with another prompt and praying the model gets it right.

Researchers introduced PSDesigner, an automated graphic design system built to solve this exact bottleneck. Instead of rendering a flattened bitmap, it turns text prompts into structured, multi-layered Photoshop files.

Working like a human design studio

Standard diffusion models try to paint everything at once. Real graphic designers never work that way.

PSDesigner breaks the task down using an agentic framework that emulates a professional creative process across three specialized components.

First, an Asset Collector Agent scans the user instruction and gathers theme-related visual assets.

Next, the Graphic Planner steps in to figure out layout, composition, and visual hierarchy.

Finally, the Tool Executor carries out the actual graphic manipulations, placing elements and executing tool calls on the canvas.

The real trick is the feedback loop. The intermediate result flows back to the Graphic Planner to evaluate visual balance, sends new edit instructions to the Tool Executor, and repeats the cycle until the poster looks right.

Learning from traces, not just finished files

To give the system genuine tool-use capability, the researchers created a dedicated dataset called CreativePSD.

Instead of just collecting finished designs, CreativePSD packs a large volume of PSD files annotated with complete operational traces across diverse scenarios and artistic styles.

The models do not just look at a finished canvas. They learn the exact sequence of expert procedures, such as how elements are introduced, grouped, and refined over time.

In benchmark evaluations, PSDesigner outperformed existing design methods across diverse graphic tasks, successfully producing clean, layer-separated files ready for manual touch-ups.

Why it matters

We think PSDesigner takes a very clever angle: it stops obsessing over one-shot pixel perfection and starts respecting the reality of production pipelines.

For creators and non-specialists, it lowers the barrier to producing decent promotional graphics without stripping away final control. When an output lands as a PSD, swapping out a price tag or hiding an awkward visual element takes seconds in Photoshop.

For the wider industry, this addresses the fundamental flaw of existing text-to-image models in commercial settings. Flat outputs are dead ends for collaboration; layered files fit right into existing agency and marketing workflows.

There are still limits, of course. Layout quality and asset harmony will inevitably depend on the underlying models' taste and capability. Even so, building multi-agent scaffolding around native design formats feels like the most practical path forward for generative design.