I think there is.
Look into gift-config.mrml. In gift-config.mrml, query processors are built of other query processors. The separate_normalisation algorithm makes use of it. It uses 4 query engines that are configured to use the same collection but use different features on these collections. (As I recall, the queries are processed in parallel, making use of multiprocessor architectures. Debugging that stuff was good fun.) You could configure the query processors to use 4 different collections and this should do about what you want them to do.
There would be an issue, though: The ranking method of GIFT considers the other images of one collection for scoring. As a consequence, if you run 4 queries in 4 collections, the ranking values would not be quite commensurate. However, if this did not hurt you before, knowing that won't hurt you either.