# Primary Research Sources for v0.7

The runtime spec uses these as external metric/benchmark references. The AADS/GAR-specific composite score and anti-homogenization score are project-defined experimental abstractions, not claims from these papers.

1. Dombrowski et al., *Image Generation Diversity Issues and How to Tame Them*, CVPR 2025. Introduces Image Retrieval Score (IRS) for generative diversity / coverage.
2. Ghosh et al., *GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignment*, 2023. Object co-occurrence, position, count, color.
3. Lin et al., *Evaluating Text-to-Visual Generation with Image-to-Text Generation*, 2024. VQAScore and GenAI-Bench for compositional alignment.
4. Wu et al., *Human Preference Score v2*, 2023. Human preference prediction benchmark/model.
5. Xu et al., *ImageReward*, 2023. Human preference reward model for T2I.
6. Hertz et al., *Style Aligned Image Generation via Shared Attention*, CVPR 2024. Uses CLIP for text alignment and DINO embedding similarity for set/style consistency.
7. Song et al., *DiffSim: Taming Diffusion Models for Evaluating Visual Similarity*, ICCV 2025. Diffusion-feature similarity for appearance/style consistency.
8. Pan et al., *ICE-Bench: A Unified and Comprehensive Benchmark for Image Creating and Editing*, ICCV 2025. Multi-dimensional evaluation including aesthetic, imaging quality, prompt following, source/reference consistency and controllability.
9. Shum et al., *Color Alignment in Diffusion*, CVPR 2025. Fine-grained color constraint alignment while retaining diversity.