In PCA, the face images are formed from a linear combination of the eigenfaces, using both positive/negative weights.
The "eigenfaces" arise in facial recognition problems using principal component analysis (PCA). They are the eigenvectors of the sample covariance matrix of the initial face images. As with any application of PCA, the original data vectors can be represented as a linear combination of the eigenvectors of their covariance matrix --- that is, each of the faces in the original data can be obtained as a linear combination of the eigenfaces. More importantly, we can approximate faces well by taking only a truncated set of the most important eigenfaces (i.e., the ones corresponding to the highest eigenvalues).
When we represents faces in the original data as a linear combination of the eigenfaces, the elements of this linear combination can have positive or negative coefficients. So yes, a face in the original data may be represented with a negative value associated with an eigenface (i.e., it is formed by subtracting some amount of that eigenface from other eigenfaces).
The graph in the link in your question is simply showing the principal components of data point in two dimensions on a graph. The circles in this graph represent data points (not holes) and the red and green arrows show the first and second principal components of the data. So in this graph, the red and green arrows would be the "eigenfaces" (for face images with only two pixels) and each of the data points can be represented as a linear combination of these "eigenfaces". When you represent the data points in PCA you are looking at them the data on this new set of axes. So, taking the red axis as an example, you can see that some of the points will have positive values on this axis, and some will have negative values.