Computer-aided design of advertisement product compositions
summary
Claims semi-automated creation of product compositions by letting designers steer an interactive genetic algorithm that generates layouts and applies visual effects; implementation uses DEAP for evolution, Pillow for image ops, Flask and JavaScript for the web UI and runs on AWS EC2. Data limitations force design of rules rather than pure transfer learning: 200 PSDs are huge and messy and JPEG feature extraction proves infeasible. After four design iterations a pairwise test with ~50 respondents shows about 35% chosen for computer‑aided designs.
AI-written summary · about
chapters (25)
- 0:00Master thesis project intro
- 1:02Company background
- 2:08Manual ad composition
- 3:11Data-driven challenges
- 6:26Learning from designers and computer-aided design
- 7:07Genetic algorithms in design
- 9:11Interactive GA research
- 10:39Learning design rules
- 11:43Algorithm implementation
- 13:32Selection and crossover
- 14:16Mutation strategies
- 15:44Web app demo and results
- 16:10Designer collaboration learnings
- 17:13Pairwise test results
- 18:53Next steps and future apps
- 19:17Key takeaways
- 19:43Q&A
- 20:03Input size and adaptability
- 20:36Creativity in automated design
- 21:20Pairwise testing clarification
- 22:12Real-world deployment plans
- 22:33CTR in fitness function
- 22:55Random feedback baseline
- 23:27Algorithm inspiration for designers
- 24:20Final Q&A and conclusion
related talks
45mDesign Automation in JS
18mGenerative Design
34mDesign in the times of AI
01h 00mPanel Discussion: AI Meets Design: A Dialogue
11mDesign Adjectives: A Framework for Interactive Model-Guided Exploration of Parameterized Design ...
19mAI-Powered Startup Toolkit for a Product Designer
01h 27mImage Generation with AI
52mDesign Engineering: The next era of Software Design