VLMs evolve designs: CLIP-IQA and Glicko rank aesthetics
Pairwise VLM comparisons beat traditional aesthetic scoring for generative design.
Deep Dive
Researchers Krol and McCormack explored two methods using Vision-Language Models to evaluate aesthetics in evolutionary systems: CLIP-IQA scoring and pairwise comparisons with Glicko rating. In a case study with a custom generative system, they compared the resulting rankings with an artist's aesthetic ranking and other evaluation techniques, documenting the artist's experience and analyzing strengths and weaknesses of both approaches. Presented at ICCC26.
Key Points
- Two VLM methods tested: CLIP-IQA scoring and pairwise comparisons with Glicko rating system.
- Pairwise VLM with custom user prompt aligned better with artist's aesthetic rankings than CLIP-IQA.
- Published at ICCC26; offers scalable aesthetic evaluation for generative design and evolutionary art.
Why It Matters
Lets creators guide generative AI with human-like aesthetic taste, not generic metrics.