Multi-agent reasoning for authentic traditional Chinese painting synthesis

Generating Traditional Chinese Paintings (TCPs), also known as Guohua, remains challenging because such paintings are governed not only by visual appearance but also by culturally grounded principles of composition, brushwork, and symbolic intent. Contemporary text-to-image (T2I) systems can imitate ink textures, yet they often fail to preserve principles such as Liu Bai (functional blankness), Qi Yun (spirit resonance), and context-dependent Cun-fa texturing. We propose MAC-Guohua, a Multi-Agent Collaborative framework that uses reasoning-to-synthesis to translate abstract user intent into structured visual constraints. The framework integrates CP-Knowledge, a retrieval-augmented knowledge base used in a retrieval-augmented generation (RAG) workflow for Chinese painting theory, specialized agents for intent analysis, composition planning, and regional brushwork specification, and a controllable SDXL-based visual synthesizer guided by segmentation ControlNet and a LoRA adapter. To improve reproducibility, we formalize the Knowledge-to-Layout (K2L) mapping module, describe the intermediate layout representation, provide the DOI-archived CP-5K release with train, validation, and held-out test split files, release the implementation code, and evaluate the framework on disjoint training, validation, and held-out test splits. Experiments on the CP-5K dataset show improved performance over the evaluated GAN-based, SDXL-based, and ControlNet-based baselines, achieving an FID of 15.28, a CLIP score of 31.45, and a Negative Space Ratio (NSR) error of 2.1% on the held-out test set. Additional ablations, prompt-fairness baselines, convergence statistics, and latency analysis further clarify the contribution and cost of each component. Expert evaluations indicate that MAC-Guohua improves perceived compositional balance, brushwork authenticity, and cultural plausibility. The implementation code and evaluation scripts are available at https://github.com/pfeijing/mac-guohua.git , and the DOI-archived CP-5K dataset release is available at https://doi.org/10.5281/zenodo.21938347 .