Yes! The insane idea Phil Marshall and I had yesterday for finding gravitational lenses automatically worked. What we do is (a) subtract smooth models from sources in HST imaging that appear to be massive galaxies at reasonable redshift, producing residuals (including, one hopes, lensed background galaxies), (b) loop over possible parameters for a simple lens model, fixing the center of the lens at the center of the massive galaxy, (c) throw image-plane pixels back to a pixelized "source plane", but (d) taking the minimum image intensity value among the values of the image-plane pixels that hit each source-plane pixel, and (e) measuring the surface brightness inside the high-magnification (principally the multiple-imaging) part of the source plane, as a function of the lens model parameters. This method, because it takes the minimum of the image-plane pixels that hit each source-plane pixels, in effect uses the blank parts of the image plane to "veto" the possible locations on the source plane (and hence models that require that originating location) from which the image-plane intensity might originate.
The amazing thing is, it seems to work, in that it finds a high source-plane intensity (after this min-value filtering) for sources that do look like lenses, and it finds a peak in the intensity at values of lens-model parameters that are consistent with what is found from direct, manual lens modeling. Here's an example plot of the source-plane intensity as a function of the lens (singular isothermal sphere) effective radius for something in Lexi Moustakas's LEGS survey. Note the peak around 20 pixels, which Marshall found previously by modeling this source. Non-lenses show no such peak anywhere.