Research map / Image generation / Autoregressive and latent generation

From Elements to Design: A Layered Approach for Automatic Graphic Design Composition

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Layout generation · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LaDeCo, which plans semantic layers for multimodal graphic elements with GPT-4o and then fine-tunes an LMM (Llama-3.1-8B with LoRA) to predict element attributes layer by layer, feeding rendered intermediate designs back as context.

Problem

Automatic design composition from multimodal graphic elements is hard: prior work covers only subtasks (layout, typography) or flattens designs, ignoring the hierarchical layered structure that human designers use.

Approach

Introduce the layered design principle into LMMs: GPT-4o assigns each element to a semantic layer (background, underlay, logo/image, text, embellishment), and the LMM predicts attributes one layer at a time while conditioning on the rendered image of previously generated layers.

Key results

• LaDeCo achieves an overall LLaVA-OV score of 8.08 versus 5.34 for FlexDM and 6.53 for GPT-4o (GT 8.35).
• On content-aware layout generation, LaDeCo gives overlap 0.0805 and underlay effectiveness closest to GT on several metrics, outperforming PosterLLaVA and PosterLlama on most metrics.
• Removing both layer planning and layered composition drops LLaVA-OV score from 8.08 to 7.23; training on combined Crello and LargeCrello raises it to 8.22.

Measured effects reported in the paper

Surprising findings

• LaDeCo outperforms task-specific layout and typography baselines without any task-specific training when run with ground-truth earlier layers.

Limitations

• Layer planning relies on GPT-4o annotation, and the layer structure is predefined with five layers.
• Filters out samples with more than 25 elements for training efficiency.
• Does not yet accept user requirements such as textual descriptions.

Experimental setup

Datasets: Crello-v4 LargeCrello

Benchmarks: Crello-v4 LargeCrello test split of Crello-v4

Metrics: LLaVA-OV scores (5 aspects) element validity overlap alignment underlay effectiveness (Undl, Unds) canvas utility occlusion text readability

Baselines: FlexDM GPT-4o PosterLLaVA PosterLlama OpenCOLE

Builds on: FlexDM OpenCOLE LLaVA LayoutPrompter

Closest papers by meaning

• Graphic Design with Large Multimodal Model (AAAI 2025 · similarity 0.82)
• Visual Layout Composer: Image-Vector Dual Diffusion Model for Design Layout Generation (CVPR 2024 · similarity 0.80)
• Detect Anything in Graphic Design: Element-Level Rewards for Autoregressive Detection (NeurIPS 2026 · similarity 0.80)
• Desigen: A Pipeline for Controllable Design Template Generation (CVPR 2024 · similarity 0.80)
• Can Vision–Language Models Assess Graphic Design Aesthetics? A Benchmark, Evaluation, and Dataset Perspective. (ICLR 2026 · similarity 0.80)
• AesthetiQ: Enhancing Graphic Layout Design via Aesthetic-Aware Preference Alignment of Multi-modal Large Language Models (CVPR 2025 · similarity 0.79)