Personalized Visual Aesthetic Alignment via Direct Preference Optimization Do?grudan Tercih Optimizasyonu ile Kişiselleştirilmiş Görsel Estetik Hizalamasi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636993
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: CLIP, direct preference optimization (DPO), pairwise preference, personalized aesthetic modeling, scale blindness, visual aesthetic alignment
- Erciyes Üniversitesi Adresli: Hayır
Özet
Visual aesthetics assessment models often focus on general crowd preferences, thereby overlooking individual and subjective tastes. This study investigates how a generic aesthetic model can be aligned with personal preferences using limited user data and proposes a framework based on Direct Preference Optimization (DPO). Due to the instabilities and "scale blindness"caused by absolute scoring systems, the problem was reformulated into a pairwise preference structure, and the DPO algorithm was integrated into a CLIP-based model. Experimental results demonstrate that the proposed method reduced the test loss from 0.6931 to 0.5140 and increased the pairwise accuracy from 60% to 78.33%. These findings confirm that fine-tuning with DPO provides accurate personal aesthetic alignment compared to generic models.