A Study on Controlling Character Deformation through AI Image Generation Model Training and Prompt Engineering

This study investigates the control of character deformation in comics and animation using image generation AI. While deformation is a complex concept shaped by culture, authorship, and style, its foundation lies in the distortion of human proportions. To establish an objective baseline, this research focuses on facial ratios and defines five deformation points from simplified to realistic. Representative images were collected from distinctive genres—wuxia, graphic novel, live action, shojo, and super-deformed—and used to construct standard reference images through prompt-based generation and quantitative validation. A LoRA model, named D-LoRA, was trained with proportion-based prompts to reproduce these deformation points. To evaluate its effectiveness, three datasets were compared: D-LoRA, prompt-only, and alternative style LoRA. Quantitative evaluation employed FaceMesh similarity, FID, KID, and CLIPScore. Results demonstrate that D-LoRA consistently outperformed other conditions, achieving higher facial ratio similarity, lower distribution divergence, and stronger text-image alignment. This study contributes by systematizing deformation analysis, establishing validated standard images, and integrating multiple quantitative metrics for objective evaluation. Limitations include the focus on facial ratios while excluding stylistic features such as hair or accessories. Future work will extend to diverse genres, ablation tests, full-body proportions, and emotional expression control.