Image Aesthetics Quantification
CLIP-based aesthetics scoring system for hotel media ranking with domain adaptation
Built a CLIP-based aesthetics scorer for automated hotel image ranking, reducing manual curation effort while improving user experience through intelligent media prioritization.
Problem
Hotel booking platforms need to display thousands of images per property, but manually curating and ranking images by aesthetic quality doesn’t scale. Existing aesthetic scoring models are trained on generic datasets and fail to capture domain-specific preferences for hospitality imagery.
Approach
Developed a multi-strategy scoring system using OpenAI’s CLIP:
- Fixed scoring: Universal aesthetic reference vectors
- Context-aware scoring: Domain-specific aesthetic alignment
- Prompt ensembling: Averaged predictions across multiple aesthetic descriptors
Key optimization: Precomputed positive/negative aesthetic centroids, reducing per-image comparisons from ~2,000 to 2 for low-latency production deployment.
Results
- Successfully deployed for hotel media ranking across Adani’s travel platform
- Domain adaptation improved correlation with human aesthetic judgments
- Integrated into content curation pipeline with sub-second inference times
Code: GitHub Repository
Write-up: Medium Article