Umělá inteligence v kódování

This master’s thesis explores the innovative concept of semantic compression of multimedia data using generative artificial intelligence (AI) models. The primary objective of this research is to verify the hypothesis of whether transforming visual information into text-based semantic instructions (prompts) and its subsequent AI reconstruction can significantly increase data transmission efficiency compared to traditional compression algorithms. The methodological core of the thesis is a zero-shot experiment applied to a dataset of 100 original, previously unpublished photographs, thereby preventing data retrieval from the models’ pre-trained memory. The thesis is structured into three dimensions: the theoretical part maps the evolution of information coding and the architecture of large language models (LLMs). This theoreti- cal section concludes with a critical discussion of the ethical and legal implications of this 4 technological leap, emphasizing the disruption of copyright paradigms. The practical part quantitatively and qualitatively tests the semantic compression and decompression process across three leading AI model families. The experiment demonstrated extreme data savings exceeding 99 %, albeit at the cost of losing the structural fidelity of the original. Consequently, the thesis provides a comprehensive evaluation of the technological potential and legislative challenges of this emerging paradigm of information transfer, and proposes future research directions.