DiffusionGPT is a text-to-image generation system based on Large Language Models (LLM). It constructs domain-specific trees for various generation models using diffusion models, enabling seamless adaptation to various types of prompts and integration of domain expert models. Moreover, DiffusionGPT introduces an advantage database where thought trees are enriched with human feedback, aligning the model selection process with human preferences. Through extensive experiments and comparisons, we demonstrate the effectiveness of DiffusionGPT and its potential to push the boundaries of image synthesis across different domains.