PlayGround
Editable graphic design
PlayGround mascot holding a paintbrush, from rectangle (1).pdf

PlayGround

Progressive Layout Generationwith Render-Grounded Planning

Paper Code BibTeX

Pipeline Overview

TL;DR

We propose PlayGround, a progressive layout generation framework that adaptively adjusts the design workflow based on the current rendered canvas. Experiments on Crello and LICA demonstrate improved design quality and zero-shot generalization, with ablations validating the effectiveness of our key components.

Abstract

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 the real-world design workflows, their high-level plans are predetermined prior to execution, limiting their adaptability as the design evolves. To address this limitation, we propose PlayGround, a progressive layout generation framework that dynamically adapts to the current rendered canvas. PlayGround establishes a progressive design loop comprising a Progressive Planner, a Layout Generator, and a Visual Verifier, guided by a Structure-Aware Retriever. Given an initial set of unordered elements, the Structure-Aware Retriever initially infers a coarse layout structure to retrieve structurally similar reference designs. Subsequently, at each turn, the Progressive Planner examines the current rendered canvas alongside the remaining elements, forms the next functional group, and plans its composition. The Layout Generator then predicts editable layout attributes, while the Visual Verifier evaluates rendered candidates and provides feedback when refinement is necessary. To train the Progressive Planner with step-wise supervision, we construct a two-stage data generation pipeline that extracts functional groups and turn-level planning trajectories from completed designs. Experiments on Crello and LICA demonstrate the effectiveness of PlayGround over existing methods. Our code and weights will be publicly released.

Data Construction

From completed designs to functional groups and group-wise turn-level planning supervision. The two-stage pipeline produces group regions, cumulative canvas states, and group-wise turn-level plans, which are used to train the Structure-Aware Retriever, Progressive Planner, and Layout Generator.

Two-stage data construction pipeline
Two-stage data construction pipeline
Stage 1 constructs functional groups and their execution order. Stage 2 derives a turn-level plan for each group from the before-and-after rendered states.
STAGE 1

Functional grouping & execution order

An MLLM annotator groups elements that should be planned and positioned together, assigns functional roles, and determines a dependency-aware execution order. Group regions and cumulative canvas states are then derived.

STAGE 2

Turn-level planning annotation

With grouping and execution order fixed, before-and-after renders provide visual evidence for each group’s placement. The annotator produces grouping, execution, and layout rationales together with a qualitative layout instruction.

PlayGround

Overall pipeline
Complete PlayGround pipeline with retrieval, progressive planning, layout generation, and visual verification
Select a module to follow the pipeline
The accepted canvas and remaining elements define the input state for the following turn.
01

Structure-Aware Retriever

Global composition guidance · Once before generation

The Retriever infers a coarse layout structure from the input elements as functional group role–region pairs. It retrieves five references using role-aware spatial similarity, canvas aspect ratio, and text–image content counts. These references remain fixed throughout generation.

Query and retrieved references
Query and retrieved references
The retrieved designs exhibit similar coarse layout structures and canvas geometry.
02

Progressive Planner

Dynamic grouping · Render-grounded composition

The Planner observes the current canvas and remaining elements, forms a functional group, and plans its composition. In the example below, it uses the available upper-left space for text, while the Static Planner, which must produce the entire plan at once and therefore imagine the intermediate canvases, targets a central placard already occupied by photographs.

Progressive versus static planning
Progressive versus static planning
The same current group, different planning context. Our Planner adapts to the rendered canvas rather than relying on a fixed initial plan.
03

Layout Generator

From a qualitative plan to editable attributes

The Layout Generator predicts element-level attributes for the current group while keeping previously placed elements unchanged. It generates multiple candidate layouts because a group-level plan can admit different rendered outcomes.

Bounding boxesGroup-local stackingText attributes
Five actual candidate renders cropped from the supplied pipeline figure

Five candidate renders from the pipeline overview.

04

Visual Verifier

Inspect rendered candidates · Accept or retry

The Verifier selects a rendered candidate and either accepts it or returns a revised layout instruction for regeneration. Up to three verification rounds are allowed; if none is accepted, it makes a final selection among the per-round finalists.

Accept
Update the canvas and remaining elements.
Retry with feedback
Revise the instruction. Regenerate the same group.

Results

Quantitative comparison

Table 1. Six VLM-based criteria (mean ± sample standard deviation) and two rule-based metrics. VLM scores range from 1 to 5; higher is better. T.Occ. (text occlusion) is the fraction of text glyph pixels hidden by elements stacked above them; Rea. (text unreadability) measures how visually busy the background beneath the text is. Lower is better for both.

Crello quantitative results from Table 1 of the supplied manuscript.
MethodVLM-based metrics ↑Rule-based metrics ↓
(i)
Composition		 (ii)
Typography		 (iii)
Elements		 (iv)
Color		 (v)
Style		 (vi)
Intention		 T.Occ. Rea.
Bold and underline mark the best and second-best means among the evaluated methods, excluding GT. VFLM places text on a fixed ground-truth background.
(i) Layout composition rationality(ii) Typography legibility and hierarchy(iii) Visible element presentation(iv) Color harmony and contrast(v) Visual style coherence(vi) Design intention alignment

Qualitative comparisons

Qualitative comparison · Crello
Qualitative comparison · Crello
From the same supplied elements. We compare rendered outputs with the corresponding ground-truth designs. Columns follow the source figure: GT, FlexDM, LaDeCo, PosterCopilot, VFLM, and PlayGround.

Ablation Study

We ablate PlayGround on a balanced subset of 600 Crello test designs. VLM denotes the mean of the six VLM-based scores; higher is better.

Table 2. Pipeline components. Highlighted cells differ from the full model (I). Static is a separately trained planner that produces the entire plan once.
#RetrieverPlannerKVerifierVLM ↑
(I) FULLTop-5✓5✓3.958
(II)✗✓5✓3.901
(III)Lowest-5✓5✓3.867
(IV)Top-5✓3✓3.923
(V)Top-5✓1✓3.891
(VI)Top-5✓5✗3.698
(VII)Top-5Static5✓3.894
  • Retrieval. Top-5 references help (I vs. II), while the lowest-ranked references are worse than none (III): mismatched references harm generation.
  • Candidates. More layout candidates steadily help: K = 1 → 3 → 5 (V → IV → I).
  • Verifier. Replacing it with random selection causes the largest drop (VI).
  • Planner. Planning once at the start (VII) underperforms progressive planning from intermediate renders.
Table 3. Per-turn granularity and grouping. Zero-shot with GPT-5.6 Luna, using only the Planner–Layout Generator loop to isolate these factors.
#Grouping schemePer-turn unitVLM ↑
(I)NoneElement-wise3.464
(II)NoneFull-layout3.351
(III)LaDeCo-styleSemantic layer3.532
(IV) OURSOursFunctional group3.578
  • Sequential decisions. Element-wise generation (I) outperforms placing all elements at once (II).
  • Grouping. Group-based variants improve further, and our functional groups (IV) outperform LaDeCo-style semantic layers (III), which use five predefined layers in a fixed order.

Extension to
Intention-to-Design

Users may only have a design intention without the visual and textual elements in advance. We extend PlayGround to this setting by prepending an element-generation stage.

GPT-5.6 Luna derives the textual elements and visual-element descriptions. Qwen-Image-2.1 generates the corresponding visual elements, which are then composed by PlayGround without modifying the core pipeline.

These examples on Crello demonstrate the modular extensibility of PlayGround as a downstream composition module, from element-to-design to intention-to-design generation.

Design intention
Element generation
PlayGround
Editable design
Intention-to-design · Qualitative results
Intention-to-design · Qualitative results
Qualitative examples of PlayGround applied to elements generated from a design intention.

Citation

BIBTEX · DRAFT
% Add final publication metadata before publishing.
@misc{lee2026playgroundprogressivelayoutgeneration,
      title={PLAYGROUND: Progressive Layout Generation with Render-Grounded Planning}, 
      author={Jaeho Lee and Junhwan Heo and Junghyun Park and WonJun Moon and Eunju Yang and Seungho Jang and Seongchan Kim and Seungryong Kim},
      year={2026},
      eprint={2610.18865},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.18865}, 
}

Author names are provided above. Final publication metadata will be added when available; this citation remains a draft.