This master’s thesis addresses the challenge of story visualization
using generative models, focusing on improving character consistency
and scene composition. We developed a two-phase system. The first phase
transforms text into structured scene descriptions and character profiles. The
second phase performs visualization with the FLUX.1-dev diffusion model,
supported by LoRA for character consistency, MV-Adapter for multi-view
image generation, Grounding DINO for detection, and SAM for segmentation.
The system was evaluated on the PororoSV dataset and the Cinderella
story through human evaluation of visual quality, consistency, relevance,
and background metrics. Results show that our approach outperforms
the Storydall-e model on PororoSV dataset, particularly in quality and consistency,
but is computationally intensive and limited for more than four
characters per scene. The method enhances control over visualization but
requires optimization for practical use and further research to automate the
system.
|
