FEDNet:A feature-enhanced diffusion network for efficient and universal texture synthesisOA

Texture synthesis remains challenging due to the complex and varied nature of texture characteristics.To address this issue,this paper introduces a featureenhanced diffusion network,FEDNet,which achieves both fidelity and diversity through two core technical innovations:(i)a frequency-aware residual block which enhances feature extraction during the down-sampling step of the UNet architecture,and(ii)a feature fusion connection process which integrates spatial histogram layers into skip connections.By effectively capturing the global structure and intrinsic texture attributes of the input example,FEDNet can expand the input and its sub blocks after training.Additionally,our model supports continuous texture extension via a cascading mechanism,enhancing its practical applicability.Extensive experiments demonstrate our method’s improvements over state-of-the-art approaches in both stationary and non-stationary texture synthesis tasks.