A post-processing technique for fast denoising of diffusion-weighted MRI (dMRI) images is introduced, leveraging the intrinsic redundancy in dMRI using universal properties of the eigenspectrum of random covariance matrices. This method removes noise-only principal components, enhancing the signal-to-noise ratio (SNR) and improving the quality of parameter maps for visual, quantitative, and statistical analysis. By analyzing residual statistics, the technique suppresses local signal fluctuations caused solely by thermal noise, while preserving anatomical details. It also improves the precision of diffusion parameter and fiber orientation estimation without compromising accuracy or spatial resolution. The method uses the Marchenko-Pastur distribution, derived from random matrix theory, to determine an objective threshold for principal component analysis (PCA) denoising. This threshold is based on the noise level, allowing for the removal of noise-only components. The technique preserves local signal fluctuations from various sources, including anatomical details, and offers the opportunity to correct denoised signals for Rician or noncentral-χ distributed noise bias using the method of moments. The algorithm estimates the noise level and the number of significant signal components simultaneously, using the mean of the lowest eigenvalues and the Marchenko-Pastur distribution. It then nullifies the noise-only eigenvalues and reconstructs the signal. The method is evaluated on simulated and clinical data, showing improved accuracy and precision compared to state-of-the-art techniques like adaptive non-local means (ANLM) and second-order total generalized variation (TGV). The technique is applied to various diffusion encoding protocols and demonstrates enhanced SNR and signal preservation. The results show that MPPCA outperforms ANLM and TGV in preserving anatomical details and reducing residual noise. The method is efficient, with reduced computational time compared to previous approaches, and is publicly available as part of the open-source MRtrix framework. The technique is also effective in reducing noise variance, which is crucial for accurate diffusion parameter estimation. The study highlights the importance of targeted post-processing for improving the accuracy and specificity of diffusion MRI analysis.A post-processing technique for fast denoising of diffusion-weighted MRI (dMRI) images is introduced, leveraging the intrinsic redundancy in dMRI using universal properties of the eigenspectrum of random covariance matrices. This method removes noise-only principal components, enhancing the signal-to-noise ratio (SNR) and improving the quality of parameter maps for visual, quantitative, and statistical analysis. By analyzing residual statistics, the technique suppresses local signal fluctuations caused solely by thermal noise, while preserving anatomical details. It also improves the precision of diffusion parameter and fiber orientation estimation without compromising accuracy or spatial resolution. The method uses the Marchenko-Pastur distribution, derived from random matrix theory, to determine an objective threshold for principal component analysis (PCA) denoising. This threshold is based on the noise level, allowing for the removal of noise-only components. The technique preserves local signal fluctuations from various sources, including anatomical details, and offers the opportunity to correct denoised signals for Rician or noncentral-χ distributed noise bias using the method of moments. The algorithm estimates the noise level and the number of significant signal components simultaneously, using the mean of the lowest eigenvalues and the Marchenko-Pastur distribution. It then nullifies the noise-only eigenvalues and reconstructs the signal. The method is evaluated on simulated and clinical data, showing improved accuracy and precision compared to state-of-the-art techniques like adaptive non-local means (ANLM) and second-order total generalized variation (TGV). The technique is applied to various diffusion encoding protocols and demonstrates enhanced SNR and signal preservation. The results show that MPPCA outperforms ANLM and TGV in preserving anatomical details and reducing residual noise. The method is efficient, with reduced computational time compared to previous approaches, and is publicly available as part of the open-source MRtrix framework. The technique is also effective in reducing noise variance, which is crucial for accurate diffusion parameter estimation. The study highlights the importance of targeted post-processing for improving the accuracy and specificity of diffusion MRI analysis.