BROWSE
Detail
Conference Paper
Adaptive Rule-Based Captioning for Senior-Specialized Text-to-Image Generation: A Case Study on the KS-KEIT Dataset
Cited - time in 
Share 
Share
- Authors
- Ah Reum Oh, Wooseok Song, Choi JoongYong, Hyunjoo Kim
- Issue Date
- 2026-10
- Citation
- International Conference on Consumer Electronics (ICCE) 2026 : Asia, pp.1085-1089
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Captioning large image datasets for text-to-image fine-tuning is costly, while existing human annotations are not necessarily reliable. We address both challenges through rule-based, VLM-driven captioning for KS-KEIT, a Korean senior-specialized image dataset. Using a Chain-of-Visual-Thought (CoVT) model, we investigate how category-name grounding affects caption quality and find that its effectiveness depends on the interaction between visual structure and naming specificity. Based on this finding, we derive an adaptive per-category captioning policy and apply it to the dataset. A parallel audit of existing human annotations reveals systematic cases of content corruption and context-insensitive templated captions, indicating that automated captioning can serve not only as a scalable substitute but also as a corrective mechanism for unreliable annotations. In downstream evaluation, both CoVT-captioned models outperform human-caption fine-tuning in 95% of the evaluated control categories. The adaptive strategy achieves the highest mean CLIP alignment (0.3551), providing a modest additional gain over uniform grounding (0.3531). The human-captioned model also performs below the non-fine-tuned base model on average, indicating that human-caption fine-tuning did not improve alignment in this evaluation.
- KSP Keywords
- Adaptive rule, Case studies, Image generation, adaptive strategy, fine-tuning, image datasets, parallel audit, rule based