This paper explores the potential of neural text generative models to produce poetic lines that can serve as seeds for artistic inspiration. Drawing on theories of aesthetic perception and creativity, we hypothesize that lines characterized by indeterminacy and occupying an intermediate space between randomness and predictability are most effective at inspiring creative work. We compare two architectures: Long Short-Term Memory Variational Autoencoders (LSTM-VAE) and Transformer-based Large Language Models (LLMs), analyzing their outputs using metrics that capture linguistic, stylistic, and poetic qualities. Our analysis reveals that LSTM-VAE generates lines with higher global entropy and more variable syntactic patterns while maintaining lower levels of pretentiousness compared to LLMs. While LLM-generated lines demonstrate richer conventional poetic imagery, they often present overly finished forms that may inhibit rather than stimulate creative exploration. The study’s findings suggest that smaller, more focused models trained on curated datasets might be more effective at generating semantically open lines than larger, general-purpose language models. This is particularly relevant during the Seed phase of creativity, where the goal is not to produce polished artistic output but to help artists enter a state of heightened perception and creative possibility. Our work contributes to understanding how computational systems can effectively support human creativity while providing a framework for evaluating AI systems designed to bring artists into the creative state.