VERSION 3:

So this version, attempt 39, has a larger dataset and was trained for longer; around 7500 steps out of a total of 10k steps. In general, it's improved, but there are still strengths and weaknesses.

Strengths: so the ability to change body shapes is very much improved. You should now find it much easier to go between, say, puffed arms, and sunken hands. You should also get proper 'size' changes between tight clothes and undersized clothes. You should also find it easier to pose laying down than before, in various ways. It should work well with most style loras that I tested, and some character loras.

Weaknesses: the lying tags, particularly on back, are prone to weird generation errors at times. Inpainting is almost always a must. Character loras are very hit and miss; the reason seems to be due to the character loras themselves rather than this lora though, because they didn't realize that they trained in the character's physical dimensions with the lora. As a result, some character loras really struggle to adhere to the arm and leg size tags. I'm not sure I can fix this, but I'm working on it. Torn clothes needs some work; because the dataset for the base illustrious model is sparse on this tag, it can take some work to get what you want. I feel I can probably improve that in the future.

VERSION 2:

So this is my 33rd attempt at making this concept.

This one has a bunch of changes to it. First, it has a different trigger word. It has a bunch of new words it's trained on, so there's a difference between puffed arms and sunken hands, for example. It should also be easier to get poses you want using sitting or standing as tags, and it should be mostly compatible with a lot of other LORAs, particularly style LORAs. There is also meant to be a difference between tight clothing and undersized clothing, but more on that in a moment. You can also also do purple skin, not just blue. You can also change the shape of the character by using breast or belly tags.

Problems and Challenges: There are a few things that it still struggles with; first and foremost, character loras can be used, but it often clashes with the fact that those LORAs have unknowingly trained in the body shape of those characters, which clashes with the body shape trained in this. Some tags may need to have their weights increased. While standing and sitting work fine, lying on back and front can create some real issues. Inpainting is basically a requirement to fix errors, or something like Adetailer, but relying on Adetailer may be hard as the change in shape makes those models bad at recognizing things. The thing I am most frustrated with though is the shape dynamics; I've yet to find a way to make clothing look as small as it should be, as the 'undersized clothes' tag seems woefully unable to grasp what it should be generating based on the training data. This model is far from perfect, basically. However, I do think it's an improvement from V1.

VERSION 1:

So this is my 24th attempt at training this concept using various datasets, and it's the closest I've ever come to having it be 'okay.' I'd like it to be better but I think I've hit my 'wall' with it. Overall, it does get the shape mostly right; it struggles a bit with proportions in some cases and poses. I intended for this to not be a model for the in between; I specifically wanted to train a lora that focused on the spherical shape, and it turns out it's really hard to get that right.

But, since no one else had made one, I figured I'd throw my hat into the ring, see if anyone else can build on it. Maybe I'll give it another go one day.