Open chrypnotoad opened 1 month ago
I think the current model is built upon fine-tuning on a large amount of drawing sequence data. I guess lllyasviel has this kind of data from the artists. But the data of Loomis method might be insufficient for fine-tuning (he has to recruit specific artists to draw the corresponding sequences).
I think we need way way more data to achieve that (which looks cool). I have physical copies of Loomis figure drawing and head one but I think even scanning those probably the data is not enough to train a project like that. In fact probably with sequencial imagery will be harder now that I think on it a bit more. Another option is to mix drawing methods of "blocking" but won't be a unique Andrew's representation. We could add Jack Hamm but results probably will end up mixed.
I'm very curious at how this was trained.
This could be extended to actual help learn how to draw if you could undo to a specific method like just planes, loomis head, perspective lines ect.