Image simulation plays an important role in remote sensing system design and data processing algorithm development, supposing that the fidelity of the simulated images is high enough. Many remote sensing image simulation models generate the geometric characteristics of the images through a georeferencing, convolution, and resampling process. In the georeferencing and resampling steps, each pixel is taken as a point, meanwhile a shift-invariant detector point spread function (PSF) is used in the convolution step. It omits the footprint size variation caused by the ground relief, earth curvature, and oblique viewing. To improve the fidelity of the simulated images, a pixel-size-varying (PSV) method was proposed: the four corners of each detector in a whiskbroom, pushbroom, or staring imaging sensor are separately considered in the georeferencing step, the sensor detector PSF is abandoned from the convolution step, and then the PSV sampling is simulated using an overlapping-area-weighted sum of the oversampled pixels. A validation experiment was conducted in simulating EO-1 Hyperion L1R data from georeferenced HyMap reflectance data. It showed that the PSV method outperforms the traditional method in the spectral aspect and is equal to the traditional method in other aspects, by comparing the simulated images with the actual one.