Aims. Our aims are to determine flux densities and their photometric accuracy for a set of seventeen stars that range in flux from intermediately bright ( 2.5 Jy) to faint ( 5 mJy) in the far-infrared (FIR). We also aim to derive signal-to-noise dependence with flux and time, and compare the results with predictions from the Herschel exposure-time calculation tool. Methods. We obtain aperture photometry from Herschel-PACS high-pass-filtered scan maps and chop/nod observations of the faint stars. The issues of detection limits and sky confusion noise are addressed by comparison of the field-of-view at different wavelengths, by multi-aperture photometry, by special processing of the maps to preserve extended emission, and with the help of large-scale absolute sky brightness maps from AKARI. This photometry is compared with flux-density predictions based on photospheric models for these stars. We obtain a robust noise estimate by fitting the flux distribution per map pixel histogram for the area around the stars, scaling it for the applied aperture size and correcting for noise correlation.Results. For 15 stars we obtain reliable photometry in at least one PACS filter, and for 11 stars we achieve this in all three PACS filters (70, 100, 160 µm). Faintest fluxes, for which the photometry still has good quality, are about 10 -20 mJy with scan map photometry. The photometry of seven stars is consistent with models or flux predictions for pure photospheric emission, making them good primary standard candidates. Two stars exhibit source-intrinsic far-infrared excess: β Gem (Pollux), being the host star of a confirmed Jupiter-size exoplanet, due to emission of an associated dust disk, and η Dra due to dust emission in a binary system with a K1 dwarf. The investigation of the 160 µm sky background and environment of four sources reveals significant sky confusion prohibiting the determination of an accurate stellar flux at this wavelength. As a good model approximation, for nine stars we obtain scaling factors of the continuum flux models of four PACS fiducial standards with the same or quite similar spectral type. We can verify a linear dependence of signal-to-noise ratio (S/N) with flux and with square root of time over significant ranges. At 160 µm the latter relation is, however, affected by confusion noise. Conclusions. The PACS faint star sample has allowed a comprehensive sensitivity assessment of the PACS photometer. Accurate photometry allows us to establish a set of five FIR primary standard candidates, namely α Ari, ε Lep, ω Cap, HD 41047 and 42 Dra, which are 2 -20 times fainter than the faintest PACS fiducial standard (γ Dra) with absolute accuracy of <6%. For three of these primary standard candidates, essential stellar parameters are known, meaning that a dedicated flux model code may be run.