MLDiff: a VAE-integrated diffusion model with fourier filter for high quality line drawing generation of ancient murals

Abstract

Ancient murals (e.g., Dunhuang murals) are vital cultural heritage with unique artistry and historical value, and their digital line drawings support preservation and restoration. Existing methods suffer from discontinuous edges, detail loss, or high complexity. To address the long-standing challenge of generating accurate and clear mural line drawings, we propose a novel approach rooted in the denosing diffusion probabilistic model (DDPM), drawing on the decoupled diffusion paradigm and latent-space optimization strategies tailored for line drawing generation tasks. Our method first integrates a variational autoencoder (VAE) with a decoupled diffusion model. Then, to further facilitate the separation of edge features from noise, we incorporate a specially designed Fourier filter block. Finally, a joint loss function guides the model’s training to achieve accurate, crisp, and structurally consistent mural line drawings. Extensive experimental results that the proposed method not only has certain advantages in qualitative evaluation, and the method can effectively generate accurate mural line drawings, but also exhibits good performance in quantitative model evaluation, surpassing the comparison methods in several key metrics.