Master thesis : Generative AI methods to create comic strips

Abstract

The recent surge in interest surrounding generative AI, particularly stable diffusion, highlights their transformative potential in creating realistic content across various domains, from images to text and music. These advancements promise to revolutionize content generation, opening up new creative possibilities. However, challenges persist, notably in ensuring high image quality and consistency.

The challenges addressed include generating minimally pixelated, high-resolution images swiftly, and maintaining consistency across characters, scenes, and style within comic panels. Achieving 100% consistency in stable diffusion remains elusive due to the inherent randomness in AI models trained on diverse datasets.

The research aims to create a tool enabling individuals with limited drawing skills to produce comics using generative AI. Key findings are presented, starting with an exploration of generative AI and stable diffusion, comparing older models with newer SDXL1.0 models, and selecting ComfyUI as the ideal user interface. The study delves into workflows, testing image generation for text-to-image, text-to-image with ControlNet, inpainting, and maintaining consistency through effective prompts.

The research explores alternative solutions, focusing on LoRAs for fine-tuning models and achieving both consistency and flexibility. Tests reveal LoRAs' potential in altering character appearances, generating cartoon-style images, and providing conclusive results for prompt-driven modifications. The integration of LoRAs into ComfyUI is discussed.

In conclusion, the research successfully achieves its primary objectives, showcasing the tool's capability to generate consistent, high-quality comic panels. Despite challenges, the findings contribute to advancing generative AI applications. The implications extend to potential uses in the creative industry, emphasizing the tool's adaptability and user-friendly nature.