Knowledge-constrained image generation and early-stage design screening based on Yaozhou Kiln ceramic cultural genes

Digital inheritance of ceramic heritage requires generative systems that preserve relationships among motifs, glazes, craft traces and cultural meanings. We developed a knowledge-graph (KG) - and reinforcement-learning-assisted framework for constrained image generation and early-stage design screening using Yaozhou Kiln ceramics. Grounded coding, Kano analysis and the Analytic Hierarchy Process identified cultural features, which were organised as a curated knowledge-graph subset for rule-based caption, prompt and negative-constraint retrieval. Qwen-Image was adapted using Low-rank Hadamard Product (LoHa) , while lightweight reinforcement learning optimised prompt policy and output ranking without updating image-model weights. Sequential main-process comparisons (P0), controlled ablation tests (P1), reliability and robustness audits (P2), and held-out validation (P3) separated descriptive trends from evidence derived from controls, audits and held-out tests.In P0, the KG-guided LoHa condition with reinforcement-learning feedback (G6) increased PickScore from 20.860 ± 0.174 to 21.744 ± 0.119 and ImageReward from 0.419 ± 0.038 to 0.739 ± 0.028 relative to baseline; Mean Expert Score increased from 3.12 ± 0.10 to 3.95 ± 0.10. G6 performed best on preference-related indicators, whereas the KG-guided LoHa condition without preference feedback (G5) achieved the lowest Fréchet inception distance (FID). KG-guided retrieval remained higher than the tested prompt-length and random-label controls. Together, these findings support an auditable framework for culturally constrained image generation and expert-led early-stage design screening under the evaluated conditions.