Some problems in reproduction

[Jungle12138]

Thank you for your wonderful work! I recently ran into some questions replicating your model on my own dataset and would like your answers.

1. Limited by the fact that my GPU only has 16g of video memory, the GPU memory usage in the test stage reached 20G+ (borrowed part of the system memory), while the GPU memory usage in the training stage was less than 10G(did not borrow the system memory), is this situation normal?
2. Since NFE=1,2,3,4,5 pictures in total will be generated for each picture in the test stage, which is very time-consuming, may I ask where to modify the code to generate only pictures such as NFE=5? I am looking forward to your answer. Thank you