@KohakuBlueleaf can you have a look at it?
Should a remove the pbar_gui bits as they seem to be getting merged from #8668 already?

@KohakuBlueleaf can you have a look at it? Should a remove the pbar_gui bits as they seem to be getting merged from #8668 already?

Yeah I think this PR should be for crop/pad only
pbar should be discussed in #8668 or another PR

@nocrcl I did some review and I have some opinions on it not sure if you want to apply it

Basically for this kind of more comprehensive preprocess things I may consider a standalone node for it.
So possible we modify the dataset loading to directly return original images than we setup different preprocess node for different preprocess mode or method.

Which make the code more clean and have more potential in complex preprocess pipeline.

cc @comfyanonymous, not sure how you think about it. I think introduce the possibility of "workflow for dataset processing" is actually a good idea and can be a standalone feature as well.

@KohakuBlueleaf That sounds great.
I actually tried this before, to load the image set and apply image processing inside comfyui, but it wouldn't load images with different sizes (which makes sense for the training part).

Still, I love the simplicety and the possibility to get complex with this node set! The way the crop and pad works in this PR made it possible for me to smash a few images into a folder and get a LoRA trained in ComfyUI without thinking too much about the resolutions of the source images.

I may just not be the right one to pull this off. My knowledge and time are just... meh.

@KohakuBlueleaf That sounds great. I actually tried this before, to load the image set and apply image processing inside comfyui, but it wouldn't load images with different sizes (which makes sense for the training part).

Still, I love the simplicety and the possibility to get complex with this node set! The way the crop and pad works in this PR made it possible for me to smash a few images into a folder and get a LoRA trained in ComfyUI without thinking too much about the resolutions of the source images.

I may just not be the right one to pull this off. My knowledge and time are just... meh.

Unfortunately put those processes into the load image function actually introducing more complexity in our view.

But we have discussed about the dataset process nodes thing and I will implement them.