A Review of Artistic Image Generation and Evaluation: Models, Metrics, and Applications

• Hanhoon Kang,
• Sung-Dong Moon,

Vol. 20, No. 1, January 31, 2026
10.3837/tiis.2026.01.002, Download Paper (Free):

• Image Generation AI
• Evaluation Metrics
• Artistic Images
• stable diffusion
• LLaVA

Abstract

We review recent advances in text-to-image (T2I) generation with emphasis on artistic image synthesis and targeted discussion of anatomical use cases to expose evaluation gaps. Empirically, state-of-the-art systems (Stable Diffusion, DALL·E, MidJourney, LLaVA) reliably reproduce impressionist/abstract styles but exhibit degradation in realism and anatomical fidelity, including geometric distortion and semantic oversimplification. We find that pixel/distribution metrics (PSNR, SSIM, FID) insufficiently track perceived artistic quality, while semantic metrics (CLIP, GLIPS, ArtScore, VQA) provide partial signal yet diverge from expert judgments. We propose a hybrid evaluation framework that (i) couples automated metrics with domain-expert assessment, (ii) reports uncertainty and inter-rater agreement, and (iii) uses curated datasets spanning art movements and synthetic anatomical imagery. This framework yields more stable rankings across styles and tasks and clarifies failure modes relevant to education and diagnosis. We conclude with actionable guidelines— balanced metric suites, task-specific benchmarks, and transparent reporting—to advance reliability, safety, and controllability of T2I systems at the intersection of art and medicine.

Statistics

Cite this article