Whitened Score diffusion models enable stable training with arbitrary Gaussian noise by avoiding covariance inversion, improving performance on inverse problems with structured noise.