pattern design needs breadth without chaos and repeat safety, yet current generators blur structure and drift in colour.This paper introduces a hybrid generator that separates structure from surface using a latent scaffold refined by diffusion and guided by differentiable priors.In the scheme, first a variational encoder learns a compact scaffold for repetition, alignment, and palette allocation; then a latent denoiser sharpens edges and adds micro variation; finally lightweight priors keep symmetry on beat, close borders for seamless repeats, and hold the palette steady with simple knobs.Experiments on two curated corpora show structural coherence rises by 11%, seam energy drops by 51%, and palette error falls by 26% against strong baselines, while sampling stays about half a second at 512 pixels and human preference reaches 71%.
Gu Gong (Thu,) studied this question.