A new method for optimal image subtraction is presented, designed to subtract two images with different seeing. Image subtraction is essential for analyzing microlensing survey images, but perfect subtraction is difficult due to the need for an accurate convolution kernel. The authors demonstrate that an optimal kernel can be derived using a least squares analysis of all pixels in both images, and that differential background variation can also be fitted simultaneously. The method also handles PSF variations effectively. The method was tested on OGLE II data from a Galactic Bulge field, and the results showed residuals close to photon noise expectations. Light curves of variable stars were also analyzed, showing error distributions close to those expected from photon noise. The method is shown to be effective even in crowded fields, and is suggested to be particularly useful for microlensing surveys, where it could significantly improve photometric accuracy and completeness. The method involves registering frames to a reference, aligning seeing, and solving for the kernel using a least squares approach with a basis of functions. The method also accounts for PSF variations and differential background subtraction. The method was applied to OGLE data, and the results showed that the residuals were close to Poisson statistics. The method is efficient, with a computing time of about 1 minute for a 1024x1024 frame on a 200 MHz PC. The authors acknowledge the OGLE team for providing the data and thank several individuals for their support.A new method for optimal image subtraction is presented, designed to subtract two images with different seeing. Image subtraction is essential for analyzing microlensing survey images, but perfect subtraction is difficult due to the need for an accurate convolution kernel. The authors demonstrate that an optimal kernel can be derived using a least squares analysis of all pixels in both images, and that differential background variation can also be fitted simultaneously. The method also handles PSF variations effectively. The method was tested on OGLE II data from a Galactic Bulge field, and the results showed residuals close to photon noise expectations. Light curves of variable stars were also analyzed, showing error distributions close to those expected from photon noise. The method is shown to be effective even in crowded fields, and is suggested to be particularly useful for microlensing surveys, where it could significantly improve photometric accuracy and completeness. The method involves registering frames to a reference, aligning seeing, and solving for the kernel using a least squares approach with a basis of functions. The method also accounts for PSF variations and differential background subtraction. The method was applied to OGLE data, and the results showed that the residuals were close to Poisson statistics. The method is efficient, with a computing time of about 1 minute for a 1024x1024 frame on a 200 MHz PC. The authors acknowledge the OGLE team for providing the data and thank several individuals for their support.