Dear Marco,
The source reconstruction algorithm in SPM relies on certain assumptions that I suspect your data violates. As a result your data looks much like noise for the algorithm. I would suggest you to invert your two original conditions together rather than the difference and then export two images per subject and put them in paired t-test design. Furthermore, 20 ms is a very short window, depending on your sampling rate it can contain just a few samples. You should perhaps invert a much longer window especially if the peak you are interested in is quite prominent and then focus on your short window when summarizing your result as an image. Perhaps look at the multimodal face evoked responses tutorial in the manual and use that as the basis for your analysis.