Generally I would go for I-tasser especially for trans-membrane proteins based on CASP experiments and with low sequence similarities it's my first choice.

but in your case i think Rosseta would be having more fidelity due to RossetaAntibody algorithm.

you can view this paper, it's highly related

http://www.nature.com/nprot/journal/v12/n2/abs/nprot.2016.180.html

I think you have done that, but if you didn't this protocol would be very helpful in comparison

https://link.springer.com/protocol/10.1007%2F978-1-61779-588-6_18

I-Tasser is more reliable in my opinion for predicting protein structure 

Different tools likely lead to different results. Try both.  There are another candidates to consider, like CABS-fold or UNRES. It would be great to validate the rationale of each approach by (partial) agreement between predicted and experimental results 

Validation methodology will surely helps you in this direction.

For antibodies and antibody fragments, there are a large number of structures available as templates for homology models, as well as a number of servers and programs that specialize on antibody modelling. Anything except CDR H3  can be modelled with high precision. For long CDR3, a rule-based approach may be better than free loop modelling, however, even the structures in the pdb may be influenced by crystal contacts.

I-Tasser is the best homology modeling server I have used (compare to Phyre2, RaptorX, SwissModeler, etc...). However, if the protein is not a monomer in solution, I am not sure that I-Tasser can handle this. For example, I have modeled a protein that exists as a dimer in solution with I-Tasser, and it outputs a single chain. Let me know if you have more specific questions about I-Tasser

Based on http://cameo3d.org/ the best today is Rosetta

For generic modeling of structures with low homology to potential template structures, you are right. However, in generic modeling, you assume the most important parts are the prediction of the fold, the interactions in the core of the protein and the most conserved parts of the protein.

In contrast, when modeling antibody variable domains, the core structures and their variability between germlines is very well know (>5000 structures), however, the most interesting parts are the variable loops that form the paratope, and the source of their variability to a large part does not stem from random mutations, but from specific mechanisms of genetic recombination. In addition, the detailed influence of specific sequence variants on CDR conformation and relative domain orientation has been extensively studied. You should, at least for comparison, look at https://sysimm.ifrec.osaka-u.ac.jp/rep_builder/ for modeling the antibody variable domains.

Hello

I-Tasser is the best.