Diffusion python images

[christopher5106]

I'm wondering about this normalization during training
https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py#L1079
that means images are between 0 and 1
while image preprocessing and post processing deal with pytorch images between -1 and 1 usually ?
do you have any insight about that ?

[tolgacangoz]

Could you examine torchvision.transforms.ToTensor() and torchvision.transforms.Normalize(); you will see how they convert the data.

[tolgacangoz]

ToTensor() makes them [0, 1], and Normalize() makes them [-1, 1]. Could you elaborate on the meaning of this sentence: "that means images are between 0 and 1"?

[christopher5106]

I know, but my question is more about the usage of normalisation to -1:1 in diffusion:

https://github.com/huggingface/diffusers/blob/main/src/diffusers/image_processor.py#L152

[asomoza]

I haven't tested this and I'm not the one that implemented the code, but I can think of two reasons:

• The original implementation and training used this normalization
• Related to one also, neural networks often perform better when working with input values centered around zero.

Here is a not too long more detailed explanation of the last point that I liked.

[christopher5106]

Ah sorry I missed the point that normalise bring images to range -1,1

Thank you for your help

[asomoza]

oh so your question was just that you didn't understood the difference between vanilla pytorch and torchvision. I also misunderstood your question.