Enhancing story visualization via subject-aware inpainting with SOEDiff
Abstract
Story visualization generates keyframe images from narrative text, requiring subject consistency. Existing methods, including IR-Diffusion, often fail to render small-region subjects. We propose an inpainting-based pipeline using SOEDiff, a diffusion model fine-tuned for small-object editing. By leveraging rehearsal images and subject masks, our method reformulates story visualization as localized inpainting. Experiments demonstrate improved rendering of small-region subjects while maintaining aesthetic quality and text alignment comparable to IRDiffusion.
Fields of Study & Subjects
Generative Adversarial Networks and Image Synthesis
Multimodal Machine Learning Applications
Digital Media and Visual Art
Publication details
- Journal / source
- International Workshop on Advanced Imaging Technology (IWAIT)
- Pages
- 140–140
- Publication date
- 27 February 2026
- Language
- EN