Evaluating generative media
Generative media quality is multi-dimensional. An image can be beautiful but fail the prompt, match the prompt but violate policy, or look good at thumbnail size while containing subtle artifacts.
Metric families
| Metric | Measures | Limits |
|---|---|---|
| FID | distribution similarity to reference images | weak for individual outputs |
| CLIPScore | text-image alignment | can miss quality and safety |
| Human preference | perceived quality and usefulness | expensive and subjective |
| Task success | whether output works in the product | requires task-specific rubrics |
| Safety checks | policy, likeness, explicit content, misinformation | false positives and misses |
| Provenance checks | source, watermark, disclosure, license | metadata can be stripped |
Evaluation checklist
- Prompt adherence and required elements.
- Visual or audio quality.
- Artifact rate: hands, text, faces, background, motion, clipping.
- Diversity across seeds.
- Safety and policy compliance.
- Rights, consent, and provenance.
- Latency and cost for the target workflow.
Human eval design
Use pairwise comparison when ranking candidates, rubric scoring when judging required properties, and expert review when rights, safety, medical, legal, or brand constraints matter.
Pitfall
Do not optimize only for aesthetic preference. Product media often needs constraints: accurate object, correct brand, consistent identity, usable layout, and safe context.
Connects to: task-specific evals · human evaluation · CLIPScore limits