This paper presents a novel method for optimal image subtraction of two images with different seeing conditions, which is essential for the analysis of microlensing survey images. The method involves deriving an optimal convolution kernel using a least squares analysis of all pixels in both images, allowing for simultaneous fitting of differential background variations and handling of PSF variations. The authors demonstrate the effectiveness of the method by analyzing images from the OGLE II project, showing that the residuals in the subtracted images are close to photon noise expectations. They also present light curves of variable stars, demonstrating that the method can achieve nearly optimal differential photometry even in crowded fields. The algorithm is particularly useful for microlensing surveys, improving photometric accuracy and completeness. The paper includes detailed descriptions of the method, its application to OGLE data, and discussions on sources of noise and improvements to the reference frame. The computational efficiency of the method is also discussed, showing that it can process large images quickly.This paper presents a novel method for optimal image subtraction of two images with different seeing conditions, which is essential for the analysis of microlensing survey images. The method involves deriving an optimal convolution kernel using a least squares analysis of all pixels in both images, allowing for simultaneous fitting of differential background variations and handling of PSF variations. The authors demonstrate the effectiveness of the method by analyzing images from the OGLE II project, showing that the residuals in the subtracted images are close to photon noise expectations. They also present light curves of variable stars, demonstrating that the method can achieve nearly optimal differential photometry even in crowded fields. The algorithm is particularly useful for microlensing surveys, improving photometric accuracy and completeness. The paper includes detailed descriptions of the method, its application to OGLE data, and discussions on sources of noise and improvements to the reference frame. The computational efficiency of the method is also discussed, showing that it can process large images quickly.