Fig. 1
A pipeline of the diffusion-based inpainting scheme. (a) Synthesise of an ensemble of fBm trajectories \(\{x_i\}_{i=1}^{n}\) of a certain length n and Hurst parameter H. Each trajectory is then converted into an EDM matrix \(\{a_{ij}\}_{i,j=1}^{n}\). (b) The Denoising Diffusion Probabilistic Model (unconditional DDPM, Algorithm 1 in Methods) is trained on the ensemble of EDM matrices. (c) The trained diffusion model is further used for the inpainting of masked values in incomplete EDMs (conditional generation, Algorithms 2–5 in Methods). Two examples of the input are shown: the masked entries are randomly dispersed in the EDM matrix (top) and entire rows and columns are masked (bottom). The latter example mimics the experimental FISH data.
