About training transforms.Normalize

[crapthings]

@sayakpaul

is this same thing when training?

transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
transforms.Normalize([0.5], [0.5])
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])

what is different on result image?

[sayakpaul]

transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
This normalization is designed for 3-channel images (like RGB).

transforms.Normalize([0.5], [0.5])
This is meant for single-channel images (like grayscale).

transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
Same as first.

[crapthings]

so i see all training code use this one

transforms.Normalize([0.5], [0.5])
This is meant for single-channel images (like grayscale).

what is different on inference result between two after training with single channel vs rgb channel?

https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth.py#L664
https://github.com/huggingface/diffusers/blob/main/examples/controlnet/train_controlnet.py#L687

[crapthings]

@sayakpaul if inference code doesn't denorm will that make final image color offset?

[sayakpaul]

The inference code doesn't have to denorm no? We don't use the VAE encoder during inference. What am I missing?