@Fatbuns I agree that it is time consuming. Easiest way to do it is while reading. Text also makes some of the images pretty hard to keep for training as it sometimes gets in the way of characters. However, I have seen the emergence of a few ways to somewhat lessen this issue. For instance, text (most likely just a watermark remover which could be useful but much less than a text remover) and an ai text remover, though I cannot locate it right now, so it might be a delusion of mine, not sure.

As for watermarks, I believe Mangadex would be a good source. So long as the uploaders of the manga/manhua/manhwa do not upload their own watermark, you should be good. Either way, these tend to generally be in a place they don't bother readers. Also, there are websites that do the watermark removing (ex: https://www.watermarkremover.io/) and this website, (https://www.watermarkremover.io/) if used carefully, could potentially salvage some images.

Edit: https://www.watermarkremover.io/ works really well but I the "fill" option seems to be broken. Also, while it struggles on text bubbles overlapping a character, it seems incredibly competent at making it look as though watermarks, sfx, and other small unsavory details (when it comes to images being good for training LoRAs, for instance, like movement lines, were never there in the first place. I really suggest it.