The most important thing is to use subsampled PSFs. The undersampling of the WPFC2 cameras results in the observed PSF being highly dependent on positioning at the subpixel level. A subsampled model should be generated, shifted to match the proper positioning of the observed PSF, and then rebinned to detector sampling. I typically subsample by a factor of 5 and interpolate the model to shift it.

An important step is to convolve the model with the pixel scattering kernel AFTER rebinning to normal sampling. Tiny Tim doesn't apply the kernel to subsampled PSFs. If it is not used, then the core of the model will be too sharp.

Guesses at PSF normalization should ignore the diffraction spikes from the spiders. The ripple patterns in the spiders are extremely sensitive to a wide variety of factors.

The PSF models will never be perfect matches to the observed ones, and they will likely be too symmetrical. This is due field- and filter-dependent aberrations which are not included in the models because they are not well characterized.

Despite these flaws, Tiny Tim models are superior to observed ones for core photometry and subtraction. Because of undersampling, interpolating an observed PSF to match the position of another doesn't work well.

A Real World Example

XZ Tauri is a binary star separated by 0.3 arcsec, with the two components differing by 2 magnitudes. It was imaged on the WFPC2 Planetary Camera (0.0455"/pixel) towards the edge of the field in F814W.

A 5x subsampled model was generated for the XZ Tau coordinates (the same PSF was used for both stars). I wrote an iterative, least-squares program which shifted the models via interpolation, rebinned and scatter-kernel convolved them, and then normalized them. The two stars were processed simultaneously. The result is shown in the Figure 1. The subtractions were pretty good, except there was an excess on one side of both stars, indicating that there was some coma. I adjusted the coma values in the parameter file produced by tiny1 to obtain a better match. There was a 4% difference in flux between the two types of models.

Figure 1. Observed and models images of XZ Tauri in the WFPC2 PC (0.0455"/pix).