Personally I have found little difference for the proteins I work with. The main difference is of course that alphafold is automated and homology modelling by yourself is not. Alphafold database has had a go at most if not all know proteins. In the case of doing it yourself you can make some better judgment calls or decisions with the warning that they may be a little subjective at times. The best thing about doing it yourself is that you may know more biology (chemistry or biochemistry) than alphafold. Of course for unstructured or disordered regions alphafold is no better than you are. I would suggest you try then compare with alphafold.

When alphafold was announced I assumed it was going to solve the protein folding problem ab initio and was disappointed to find it was essentially an AI automated way of doing what we do to make homology models.

Thank you Dr. Gary James Hunter. Actually, I don't know how to objectively assess the structural quality of homology models. Tools like ERRAT, PROCHECK, WHATIF, and others are too obsolete and results can be interpreted in different ways. In some cases, in case of few mutations from the reference crystal structure, even manual "substitutions" on the reference structure generated a high quality structure. I think you are right. I should apply a range of different options (AlphaFold, RosettaFold, MODELLER, SWISS_MODEL, manual structural editing based on reference structure, and others) and evaluate them using a variety of evaluation tools for stringent compliance.