iPoster: An Interactive Poster Content-Aware Layout Generation System Based on Graph-Enhanced Diffusion Model

iPoster is an innovative interactive poster design system that uses graph-enhanced diffusion models to achieve content-aware automatic layout generation. By modeling semantic connections between elements through graph neural networks and combining diffusion models to generate high-quality layouts meeting design aesthetics, this system aims to address issues like low efficiency of traditional design tools and insufficient semantic understanding of existing automatic layout methods, while supporting users to express design intentions interactively.

Poster design needs to balance aesthetics, information hierarchy, and user experience. Traditional tools rely on manual adjustments by professional designers, leading to low efficiency. Automatic layout generation is an important direction, but existing methods face three major challenges: 1) Understanding content semantic relationships to convey information accurately; 2) Meeting user intentions while maintaining design flexibility; 3) Handling complex constraints of multiple elements to generate a coordinated visual presentation.

iPoster adopts a "Graph-Enhanced Diffusion Model" architecture, which includes three main modules:

1. Content-Aware Encoding Module: Identifies types and attributes of elements such as text and images, and analyzes semantic connections (e.g., relationships between titles and body text, images and captions);
2. Graph Neural Network Modeling: Treats elements as nodes and spatial/semantic/visual dependencies as edges to capture hierarchical structures, alignment constraints, brand norms, and special user requirements;
3. Diffusion Model Generation: Generates high-quality layouts through step-by-step denoising, with advantages including high generation quality, strong controllability, and rich diversity.

The key technical features of iPoster include:

1. Interactive Intent Understanding: Supports users to specify element positions/sizes, describe styles or information focuses, and encodes inputs as constraints into the generation process;
2. Multi-Scale Layout Generation: Uses a coarse-to-fine strategy—first determining the overall structure, then refining element positions and sizes—to balance global coordination and local flexibility;
3. Responsive Design Support: Automatically adjusts layouts according to the size ratio of the target display area to adapt to different devices and scenarios.

The research team verified the effectiveness of iPoster through various poster tasks (academic conferences, commercial promotions, event announcements, etc.):

• Layout Quality: User studies and expert ratings show generated layouts meet or exceed manual designs in visual appeal and information transmission efficiency;
• Generation Efficiency: Generates multiple candidate solutions in seconds, significantly improving design efficiency;
• User Satisfaction: In usability tests, participants highly praised the system's ease of use and intent understanding ability.

iPoster has broad application prospects:

1. Design Democratization: Lowers the threshold for professional design, allowing non-professionals to create high-quality content;
2. Workflow Optimization: Provides intelligent assistance for designers, accelerating concept exploration and scheme iteration;
3. Cross-Domain Migration: The graph-enhanced diffusion model architecture can be extended to web design, APP interfaces, publication typesetting, etc.;
4. Personalized Design: Combining user preference learning and style transfer, it can realize highly personalized automatic design services in the future.

iPoster currently has limitations: 1) Its ability to understand highly abstract/avant-garde design styles needs improvement; 2) Aesthetic preferences and design norms from different cultural backgrounds require more data support; 3) Intent coordination and version management in multi-person collaboration scenarios need further research. The system will be optimized for these directions in the future.