We study two galaxies samples selected in near-ultraviolet (NUV) and in far-infrared (FIR) for which the spectral energy distributions (SEDs) from the far-UV (FUV) to the FIR are available. We compared the observed SEDs to modelled SEDs with several star formation histories (SFHs; decaying star formation rate plus burst) and dust attenuation laws (power law + 2175 Å bump). The Bayesian method allows one to estimate statistically the best parameters by comparing each observed SED to the full set of 82 800 models. We reach the conclusion that the UV dust attenuation cannot be estimated correctly from SED analysis if the FIR information is not used. The dispersion is larger than with the FIR data. The distribution is also not symmetrically distributed about zero: there is an overestimation for UV-selected galaxies and an underestimation for FIR-selected galaxies. The output from the analysis process suggests that UV-selected galaxies have attenuation laws in average similar to the Large Magellanic Cloud extinction law while FIR-selected galaxies attenuation laws resemble more the Milky Way extinction law. The dispersion about the average relation in the log (Fdust/FFUV) versus FUV-NUV diagram (once the main relation with FUV - NUV accounted for) is explained by two other parameters: the slope of the attenuation law and the instantaneous birthrate parameters b0 for UV-selected galaxies and the same ones plus the strength of the bump for the FIR-selected galaxies. We propose a recipe to estimate the UV dust attenuation for UV galaxies only (that should only be used whenever the FIR information is not available because the resulting AFUV is poorly defined with an uncertainty of about 0.32): AFUV= 1.4168 (FUV - NUV)2+ 0.3298 (NUV - I)2+ 2.1207 (FUV - NUV) + 2.7465 (NUV - I) + 5.8408.