Different inference results between OpenVINO and Torch

[FedericoDeBona]

I've trained a Padim on a custom dataset, then exported the model in both Torch and OpenVINO formats. The problem is that I get different inference results:
Torch

OpenVINO

Anomalib version: 1.2.0dev
OS: Ubuntu
Same problem with anomalib 1.1.1 installed with pip.

When using Torch, the prediction score is always better; with good images, the score is much lower than when using OpenVINO.
There is not much difference between Torch GPU vs Torch CPU. I would like to use OpenVINO on CPU, but like this is not usable. Any tip on this?

code used to train:

datamodule = MVTec(
    category=CATEGORY,
    root=f"{ROOT}/datasets_original",
    image_size=(256,256))
model = Padim()
engine = Engine(default_root_dir=RESULTS_DIR)
engine.fit(datamodule=datamodule, model=model)
engine.export(
    model=model,
    export_type=ExportType.TORCH
)
engine.export(
    model=model,
    export_type=ExportType.OPENVINO,
    compression_type=CompressionType.INT8
)
test_results = engine.test(
    model=model,
    datamodule=datamodule,
    ckpt_path=engine.trainer.checkpoint_callback.best_model_path,
)

[FedericoDeBona]

As said in #2136, if I manually resize the images to 256x256 before inference it works fine. So it's missing the auto resize in OpenVINO while Torch has it?