Four Pipelines, One Model

An image classifier was trained on photos prepared like this, in this order:

• resize to $224 \times 224$
• apply a random horizontal flip and a random brightness jitter
• divide every pixel value by $255$
• subtract $0.45$ — the average pixel value across the whole training set, measured after dividing by $255$

Training is finished and the weights are frozen. Four engineers now each write the code that will prepare photos when the model makes real predictions. All four begin by resizing to $224 \times 224$. After that, one copies the rest of the training recipe word for word; the other three drop the random steps and change the order of what is left, the numbers in it, or both.

Which of these, applied after the resize, is the correct prediction-time pipeline?