Optimising ancient costume image generation using the stable diffusion model: a focus on dynastic characteristics
Industria Textila · 30.06.2026 · 382
DOI: 10.35530/it.077.03.202570 ↗Kurz gefasst
Fachartikel in der Zeitschrift Industria Textila, online erschienen am 30.06.2026. Verfasst von JIAN HUA und 3 weiteren Autor:innen.
Abstract (Originalsprache)
The source of the generated images of ancient costumes is misplaced because the process of generating ancient costume effect images cannot accurately capture the distinctive features of different dynasties. By leveraging the Stable Diffusion model, this study organises costume characteristics across different dynasties into 163 textual prompts, drawing on historical literature and classical scrolls. By matching these prompts with image feature vectors, a new token embedding layer V* is introduced, which is optimised together with the cross-attention layer parameters Wk and Wv. Then, the model was fine-tuned using the Low-Rank Adaptation (LoRA) model to reduce training costs while maintaining historical fidelity. The results demonstrate that the optimised model can generate costume images that align with the corresponding dynastic and ethnic characteristics based on textual prompts. Validation experiments across the Tang, Song, and Ming dynasties show that the model achieves relatively low Kernel Inception Distance (KID) and Maximum Mean Discrepancy (MMD) values, indicating its effectiveness in generating ancient costume images. This study not only optimises the generation of ancient costume effect images but also holds reference value for the digital preservation and protection of costume cultures from other dynasties and ethnic regions.
Zitieren
APA
HUA, J., LI, Z., CHEN, R., & CHEN, Y. (2026). Optimising ancient costume image generation using the stable diffusion model: a focus on dynastic characteristics. Industria Textila, 382. https://doi.org/10.35530/it.077.03.202570
Vancouver
HUA J, LI Z, CHEN R, CHEN Y. Optimising ancient costume image generation using the stable diffusion model: a focus on dynastic characteristics. Industria Textila. 2026:382. doi:10.35530/it.077.03.202570
BibTeX
@article{HUA2026,
author = {HUA, JIAN and LI, ZHI and CHEN, RUIHONG and CHEN, YU},
title = {Optimising ancient costume image generation using the stable diffusion model: a focus on dynastic characteristics},
journal = {Industria Textila},
year = {2026},
pages = {382},
doi = {10.35530/it.077.03.202570}
}
EUVIMED