Kinda random but if you think the size of the input really matters for the model to learn well (which frankly I’m not convinced is an issue) you could add one or two hidden layers of decreasing sizes behind the large input size layers, before you concatenate them with the smaller size ones.