If you want to use a trained model for inference, the input data must conform to the format (including size) in which the model was originally trained. Basically, I think both of your proposed solutions, 1.a) and 1.b), can be tried and tested, and as you mentioned, the data must be processed by yourself first.

However, because you are performing inference on untrained data, and the data format is different from the model training data, I think the model performance will not be as expected.

Please refer to it, thanks.