How alpha masks are handled?

[PGCRT]

With Kohya scripts, I use --alpha_masks argument to ignore masked areas, mostly watermarks or unwanted parts of the image that I don't want the training to focus on.

I would like to know if this feature is present (For HunyuanVideo training on diffusion-pipe), if it needs special configuration or if it's natively supported, Thank you

[tdrussell]

It doesn't currently support any kind of masks.

[PGCRT]

Okay, so the alpha masks will either be ignored as if they aren't there or treated as black areas, resulting in black squares or shapes during inference? Thank you for your time

[tdrussell]

I'm not sure how this feature works in Kohya sd-scripts. Are the masks separate? Or is it part of the image itself as an alpha channel? For diffusion-pipe if an image has alpha channel, the transparent parts are converted to white (as usually that is the background). But they are not masked in the loss function.