zjuqx/

ZhiHua

Zhihua is the first domain-specific vision-language model dedicated to Traditional Chinese Painting (TCP) appreciation and storytelling. It combines supervised fine-tuning (SFT) with Direct Preference Optimization (DPO) and a threshold-based Retrieval-Augmented Generation (RAG) mechanism to reduce hallucinations and produce culturally resonant narratives. The model is trained on a large-scale TCP dataset containing question-answer pairs, technique annotations, and expert-level textual analyses. Extensive experiments and human evaluations show that Zhihua generates more faithful, coherent, and aesthetically aware descriptions than existing vision-language baselines.

Zhihua accepts an image of a traditional Chinese painting and free-form text questions (optional), then outputs multi-lingual narratives that explain artistic techniques, historical context, symbolism, and aesthetic values. A three-stage training pipeline—SFT → DPO → RAG—iteratively aligns the model with expert preferences while mitigating hallucinations. Domain-oriented evaluation metrics (factual correctness, stylistic fidelity, interpretive depth) are proposed to benchmark TCP understanding.