AutoGrader: Hybrid Evaluation of Generative Product Images via Contrastive Embeddings and LLM-Generated Grading Notes

High-quality product image generation is essential for scaling commercial content creation, yet evaluating AI-generated visuals remains a costly bottleneck requiring extensive human review. This paper proposes a hybrid automated evaluation framework that combines contrastive learning-based image models with Vision-Language Model reasoning to reduce manual inspection of generated product imagery. The system filters low-quality images, generates scene-specific grading criteria via LLMs, and performs weighted evaluation for each product-background pair. Across benchmark and internal datasets, the approach achieves strong alignment with expert design judgments, improves high-quality image selection, reduces review time, and supports scalable content generation workflows.