Question for Preprocessing Input of ArcFace Model

Hi, I just want to make sure of the preprocessing method for the arcface model is the one [here](https://github.com/deepinsight/insightface/blob/4f8c54a09b7341a757657adbccfe1f8751f6548c/recognition/arcface_paddle/deploy/pdserving/web_service.py#L25) below or the one named inv_preprocess [here](https://github.com/deepinsight/insightface/blob/4f8c54a09b7341a757657adbccfe1f8751f6548c/parsing/dml_csr/utils/utils.py#L65)

Because, when i tried to use the preprocessing the 512 embedding output is always random, but when i tried to use for example facenet preprocessing input, the embedding output is similar each time its generated.

I am looking forward to hear the explanation, Thank You !

Read the preprocessing in recognition/arcface_paddle/deploy/pdserving/web_service.py and compare it with inv_preprocess in parsing/dml_csr/utils/utils.py. Reproduce the differing preprocessing paths and check whether repeated inputs produce consistent 512-dimensional embeddings. Done means identifying which preprocessing ArcFace expects and explaining the source of the inconsistent output.

Written by the indexing model from the issue text.