PLAYGROUND: Progressive Layout Generation with Render-Grounded Planning
Synthetic voice
The authors propose PlayGround, a progressive layout generation framework that dynamically adapts to the current rendered canvas by establishing a design loop of a Progressive Planner, Layout Generator, and Visual Verifier that operate over functional groups of elements at each turn. To train the Progressive Planner with step-wise supervision, they construct a two-stage data pipeline using GPT-5.6 Luna to extract functional groups and turn-level planning trajectories from completed Crello designs. The authors report that PlayGround leads the baselines on all six VLM-judged metrics on Crello and five of six on zero-shot LICA, with a 32-participant user study ranking it first in approximately 81–84% of responses.
- PlayGround establishes a progressive design loop comprising a Progressive Planner, a Layout Generator, and a Visual Verifier guided by a Structure-Aware Retriever; at each turn the Progressive Planner examines the current rendered canvas, forms the next functional group, and plans its composition.
- A two-stage data construction pipeline built on Crello using GPT-5.6 Luna extracts functional groups and turn-level planning trajectories from completed designs to provide step-wise supervision; the Retriever, Planner, and Layout Generator are fine-tuned from Qwen3-VL-8B-Instruct with LoRA, while the Visual Verifier runs untrained.
- The authors report PlayGround leads all baselines on six VLM metrics on Crello and five of six on zero-shot LICA; a 32-participant user study ranks it first in approximately 81–84% of responses, and code and weights are promised to be publicly released.