Analysis of Conditional Image Generation Methods Using Color Palettes in Animal Personification Task

With the advent of large-scale pre-trained open-source diffusion models, image generation is now easily accessible to non-researchers. However, although diffusion models as Text-to-Image (T2I) models are powerful, they are not yet ready for task-specific applications because it is difficult to perform accurate generation using only text-based guidance. In this paper, we used T2I-Adapters, an external neural network that enables conditional guidance of diffusion models to ensure identity with the original cat, in order to take on the task of "generating characters from cats," the most popular type of animal personification. We achieved better performance by modifying the structure on the adapter side to apply the pre-trained model to the task.