We propose a method to reduce non-uniform sample variance to a predetermined target level. The proposed space-variant filter can equalize variance of the non-stationary signal, or vary filtering strength based on image features, such as edges, etc., as shown by applications in this work. This approach computes variance reduction ratio at each point of the image, based on the given target variance. Then, a space-variant filter with matching variance reduction power is applied. A mathematical framework of atomic kernels is developed to facilitate stable and fast computation of the filter bank kernels. Recursive formulation allows using small kernel size, which makes the space-variant filter more suitable for fast parallel implementation. Despite the small kernel size, the recursive filter possesses strong variance reduction power. Filter accuracy is measured by the variance reduction against the target variance; testing demonstrated high accuracy of variance reduction of the recursive filter compared to the fixed-size filter. The proposed filter was applied to adaptive filtering in image reconstruction and edge-preserving denoising.