PLAYGROUND: Progressive Layout Generation with Render-Grounded Planning

Forum topic · 小凯 · 2026-10-15

Summary

PLAYGROUND is a progressive layout generation framework for creating editable graphic designs that dynamically adapts to the current rendered canvas, rather than following a predetermined high-level plan. The system builds a progressive design loop consisting of a Progressive Planner, a Layout Generator, and a Visual Verifier, guided by a Structure-Aware Retriever. Given an unordered set of visual and textual elements, the retriever first infers a rough layout structure to retrieve structurally similar reference designs. In each round, the planner inspects the rendered canvas and remaining elements to form the next functional group and plan its composition; the layout generator predicts editable layout attributes; and the visual verifier evaluates rendered candidates and provides refinement feedback when needed. To train the planner with step-wise supervision, the authors constructed a two-stage data generation pipeline that extracts functional groups and round-level planning trajectories from finished designs. Experiments on Crello and LICA demonstrate the effectiveness of PLAYGROUND over existing MLLM-based layout generation methods. The paper (arXiv:2610.18865) was authored by Jaeho Lee and colleagues.

Paper Overview

Field: Computer Vision (CV) Authors: Jaeho Lee, Junhwan Heo, Junghyun Park, WonJun Moon, Eunju Yang, Seungho Jang, Seongchan Kim, Seungryong Kim Posted: 2026-10-13 arXiv: 2610.18865

Abstract (English)

Generating editable graphic designs requires arranging visual and textual elements into a coherent composition while preserving their content and editability. Although recent MLLM-based approaches employ iterative generation and visual refinement to reflect real-world design workflows, their high-level plans are predetermined prior to execution, limiting their adaptability as the design evolves.

To address this limitation, the authors propose PLAYGROUND, a progressive layout generation framework that dynamically adapts to the current rendered canvas. PLAYGROUND establishes a progressive design loop comprising:

Progressive Planner: inspects the current rendered canvas and remaining elements to form the next functional group and plan its composition.
Layout Generator: predicts editable layout attributes.
Visual Verifier: evaluates rendered candidates and provides refinement feedback when needed.
Structure-Aware Retriever: given an initial set of unordered elements, infers a rough layout structure to retrieve structurally similar reference designs.

To train the Progressive Planner with step-wise supervision, the team built a two-stage data generation pipeline that extracts functional groups and round-level planning trajectories from completed designs.

Results

Experiments on the Crello and LICA benchmarks demonstrate PLAYGROUND's effectiveness relative to existing methods.

--- *Auto-collected on 2026-10-15*