PSDesigner: Automated Graphic Design with a Human-Like Creative Workflow

TL;DR

PSDesigner is an automated graphic-design agent that emulates the human PSD-authoring workflow: it collects theme-related assets from a user instruction, then iteratively plans and executes tool calls to integrate those assets into a layered Photoshop document and refine inferior elements. Training relies on CreativePSD, a new dataset of high-quality PSD files annotated with operation traces, where source layers are first grouped by underlying visual concept before being decomposed into raw assets, metadata, and intermediate renders. The system splits into three trained/orchestrated components — AssetCollector, GraphicPlanner (the tool-calling LLM trained on CreativePSD), and ToolExecutor — and reports SOTA on graphic-design composition benchmarks (Crello-v5 and a new copyright-free PSD benchmark with complex layer hierarchies). CVPR 2026.

Key claims

A bottom-up, group-then-asset traversal of the nested layer hierarchy is the core inductive bias: each iteration plans (①) and inserts (②) the current asset, then identifies deficiencies (③) and refines (④), mirroring how human designers operate on PSD groups [§Methodology].
CreativePSD is constructed by first grouping the layers of collected PSD files based on their underlying visual concepts, then parsing the files to extract raw assets, metadata, and intermediate renders — used as supervision for both asset integration and layer refinement subsets [§Methodology, Construction figure].
The agent is decomposed into three components: AssetCollector (theme-aware asset retrieval from user instruction), GraphicPlanner (predicts tool calls; trained on CreativePSD; runs in two design modes — incorporate new assets vs. refine inferior layers), and ToolExecutor (applies tool calls to the PSD file) [§Methodology, Architecture figure].
Outperforms existing methods on (a) end-to-end user-intent-to-design translation, (b) graphic composition given assets on Crello-v5 (simple scenarios), and (c) a copyright-free PSD benchmark with complex layer hierarchies [§Experiments].
Existing graphic-design baselines that rely on T2I models + MLLMs simplify professional workflows and “result in limited flexibility and intuitiveness” — i.e. the bottleneck is the editable-file representation, not raw pixel quality [§Abstract].

Method

PSDesigner treats PSD authoring as a multi-step tool-use loop over a structured (typed, nested) document. The CreativePSD dataset is the load-bearing ingredient: rather than learn from raw rasterized designs, the authors first cluster the layers of each PSD by visual concept (this is the “layer grouping” Dejia flagged), then parse out raw assets, layer metadata, and intermediate per-step renders — producing operation traces that can be replayed as planner-executor demonstrations. Two CreativePSD subsets target the two design modes the model later operates in (asset integration; layer refinement). At inference, AssetCollector retrieves theme-related raw assets from a user prompt, GraphicPlanner (trained on the operation traces) emits a tool call describing what to insert or refine at the current step, and ToolExecutor mutates the PSD file. The loop terminates when all assets are integrated. No model sizes, training hyperparameters, or planner backbone are disclosed in the project-page abstract.

Results

Quantitative numbers are not in the public abstract; the page reports headline rankings on three evaluations: graphic-design end-to-end, Crello-v5 graphic composition (simple), and a copyright-free PSD-benchmark composition (complex layer hierarchies), with PSDesigner reported best on each [§Experiments]. Detailed metrics and ablations presumably live in the CVPR paper, which is referenced but not linked from the project page at filing time.

Why it’s interesting

This is the rare layered-design system that treats PSD authoring as a tool-use agent problem rather than a pixel-generation problem — it sits at the intersection of three currently-active wiki threads. On the layered-image side it complements ART: Anonymous Region Transformer for Variable Multi-Layer Transparent Image Generation (ART) and LICA: A Dataset of 1.5M Layered Graphic Design Compositions (LICA): ART generates K≤50 transparent layers jointly via region-crop MMDiT, LICA ships 1.55M template-native compositions as a public dataset, and PSDesigner attacks the same artifact (multi-layer PSD with typed components) but from the editing/composition angle with operation traces rather than parallel denoising. On the tool-use agent side it is the first filed graphic-design instance — the planner-executor split (planner=LLM trained on demonstrations, executor=deterministic tool runtime) is structurally similar to Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing and MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling but with a domain-specific tool API instead of MCP/web tools. On the synthetic-training-data side, CreativePSD is in the same lineage as Action100M’s pipeline-curated labels and InfTool’s verified trajectories — the verification gate here is concept-based layer grouping before trace extraction, a recipe that has no direct analog among the filed synthetic-data papers. Same author group (Henghui Ding) as the already-filed EffectErase: Joint Video Object Removal and Insertion for High-Quality Effect Erasing.