Hello Guillaume,
As a rule of thumb, I drop a subject only if the number of TRs in a condition drops below 30 after censoring. You need about 30 TRs for decent statistical power.
The approach you are taking now - increasing the threshold to 1.5 and dropping subjects with more than a few TRs censored - I would not recommend because then you don't know what you are dealing with. The motion artifacts are spatially non-linear and are further complicated by the slice timing pattern. So, if I were you, I will keep the threshold low (0.3 to 0.5 mm should be fine) and reject a subject only if the TRs in a condition drops below 30. That way I avoid signal bleeding effects and know exactly what the issue is - statistical power. As Rick said, 10% loss may not be bad depending on where exactly in your design those TRs are lost. I have had a loss of 30% but still included the subject because the number of good TRs in each condition (including rest) was greater than 30.
Hope that helps,
Best regards,
Gaurav