Inconsistency in condition transforms across different ControlNet example scripts

I've noticed an inconsistency in the transforms applied to condition images across different ControlNet training examples. In the [flux training script](https://github.com/huggingface/diffusers/blob/b75b204a584e29ebf4e80a61be11458e9ed56e3e/examples/controlnet/train_controlnet_flux.py#L755), the condition images are explicitly normalized.

However, in the [standard ControlNet training script](https://github.com/huggingface/diffusers/blob/b75b204a584e29ebf4e80a61be11458e9ed56e3e/examples/controlnet/train_controlnet.py#L689) and across all SD-based ControlNet training pipelines, normalization is not applied.

From my understanding, we generally avoid normalizing the condition image to prevent altering the input distribution. Does this reasoning also apply to flux?

Would this hold true for general condition images, such as a content image?