→ CLOSE

GENERATIVE AI POSTER DESIGN

GENERATIVE DESIGN, GENERATIVE AI, RESEARCH PROJECT

YEAR 2024
PROGRAM Stable Diffusion, Automatic 1111,
Midjourney, Dreamstudio, DALLE3,
RunwayML, SDXL 1.0, LoRA
MENTOR Jiri Oplatek, Ted Davis,
Bérénice Serra,
Dr. Invar-Torre Hollaus
SIZE(BOOK) 297 x 210 mm

YEAR 2024
PROGRAM Stable Diffusion, Automatic 1111,
Midjourney, Dreamstudio, DALLE3, RunwayML,
SDXL 1.0, LoRA
MENTOR Jiri Oplatek, Ted Davis,
Bérénice Serra,
Dr. Invar-Torre Hollaus
SIZE(BOOK) 297 x 210 mm

This project explores a methodology for integrating AI into contemporary poster design by leveraging a dataset of approximately 72,000 digital posters from the Basel Poster Collection. By analyzing posters' layered structures—comprising images and typography—the study employs AI-driven processes such as OCR masking for text detection, layer segmentation using OpenCV and OCR-SAM, and inpainting for image reconstruction. Stable Diffusion XL and LoRA models, fine-tuned on categorized posters, generate image layers aligned with specific design purposes, while ResNet50 and Hough Transform analyze composition to suggest appropriate typography.

This research demonstrates AI’s potential to enhance the poster design process by generating contextual image layers and recommending typography based on compositional analysis, positioning AI as an active collaborator in creative workflows.

This project has been selected for presentation at ATypI 2025 in Copenhagen.

Click here to see detailed process.

This research includes one preliminary experiment and three primary experimental steps.

Preliminary Experiment - Assessing Gen AI’s Potential and Limitations in Poster Design

These preliminary experiments assessed AI’s ability to describe, reproduce, and generate posters.

AI tools such as DALL·E 3, Midjourney, and Imagetoprompt.com were tested for their ability to describe poster designs based on visual elements.
A model trained on a limited dataset using RunwayML attempted to replicate Emil Ruder’s typographic posters.
A StyleGAN3-T model was trained on over a thousand posters to generate new designs.

While AI demonstrated potential in recognizing visual elements and generating stylistic variations, it struggled with accurately reproducing typography, maintaining structured layouts, and applying grid systems. These findings highlight the need for more refined training approaches, particularly by separating typography from image generation, to improve accuracy in typographic reproduction and structured layout composition.

❶ Evaluating AI’s Ability to Interpret Graphic Design Elements Through AI-Generated Descriptions (DALL·E 3, Midjourney, Imagetoprompt.com)

❷ Poster Reconstruction Using RunwayML Trained on 18 Posters by Emil Ruder

❸ Poster Generation Using StyleGAN3 Trained on 1,263 'Event' Posters (1920–2023) from the Basel Poster Collection

Primary Investigation - Decoding Text and Image Layers

The preliminary experiments revealed AI’s difficulty in accurately reproducing typography and maintaining structural consistency within layouts and grid systems. To overcome these limitations, this phase explored how separating typography and image layers could enhance AI’s ability to support the poster design process.

For text detection within image layers, EasyOCR and OCR-SAM were employed:
• EasyOCR, an open-source Optical Character Recognition (OCR) library, extracted text from images by identifying and recognizing typographic elements.
• OCR-SAM simultaneously performed text recognition, layer segmentation, and image inpainting to separate typography from image layers more effectively.

This separation enabled AI to better analyze typographic composition and spatial arrangement.

Original Poster

Text Detection

Poster after removing text
(Image Layer)

Typography Layer

Original Poster

Text Detection

Poster after removing text
(Image Layer)

Typography Layer

Primary Investigation - Training AI for Contextual Image Generation

With typography removed, Stable Diffusion XL 1.0 and LoRA models were trained to generate high-resolution image layers aligned with specific design purposes. Each model was fine-tuned based on categorized poster datasets, requiring approximately one day of training per category.

The goal was to enable AI to learn the distinct styles, themes, and functional purposes of different poster categories. Based on this understanding, the AI generated contextual background images (image layers) for new compositions, aligning with the visual characteristics of each category:

Event Posters (Graphic Shapes, 397)
Social Communication Posters (Illustration, 82)
Political Posters (Illustration, 85)
Food Posters (Photography, 305)
Consumer Goods Posters (Photography, 245)

Generated Poster (Event)

Generated Poster (Event)

Generated Poster (Event)

Generated Poster (Food)

Generated Poster (Food)

Generated Poster (Food)

Primary Investigation - Typography Recommendation for AI-Generated Images

To recommend appropriate typographic layers, ResNet50 and the Hough Transform were applied in combination:

ResNet50 analyzed the similarity between newly generated images and existing posters within the same category to identify suitable typography styles.
Hough Transform refined the process by examining the grid structures of both images and typography layers, ensuring structural coherence.

By leveraging these techniques, the model suggested typography that aligned with the composition and typographic conventions of existing posters.

Conclusion

This study presents a structured approach to integrating AI into poster design by addressing AI’s existing limitations in typography reproduction and structured layouts. By separating typography and image layers, fine-tuning models on categorized datasets, and analyzing composition through computational techniques, AI can evolve from a supportive tool to a creative collaborator in graphic design.

These findings provide a framework for enhancing early-stage design workflows, demonstrating AI’s potential to contribute to structured visual composition and typographic decision-making. Future research could explore refining typography placement suggestions and expanding AI’s role in typographic generation.