by using troughnouf's dataset for training instead of the synthetic one used so far, like: https://github.com/trougnouf/nind-denoise