Oh! This is super interesting to me, I've only been trying anime models

Would love to see any shared as a review!

Thanks! I use this automatic tagging extension for automatic1111's web UI https://github.com/toshiaki1729/stable-diffusion-webui-dataset-tag-editor

For the sex positions I think consistency of angle/composition is more important whereas for characters the poses are a lot more varied but focus on the specific outfit or other details. Really it's just the consistency of the thing you're training for that I think is most important.

Then removal of all the autogenerated tags that best describe the subject, and that's about it really. I prefer to have more than 150 images but it's also often hard to find enough that I think are good enough, I think the least I've used are about 80.

The boorus that allow complex searches are the best resource when it's a potentially huge number of images like the oral one to narrow those down to what I'm after, plus filtering out most of the ones with text, animation, greyscale etc.

Good luck with your own!

For the training you automatically tag all the images you put together, then remove the tags which are what you want to train it on, so for this one just stuff like anal and anal_penetration and for characters the tags that describe them.

For a sex position tag I think it's important to be consistent with the pose so that it comes out right (if you don't, you often get a garbled mess of limbs and bodies instead) but for characters, a variety of poses is good so you can capture their features from different angles.

@kwik no it's just a case of removing all the tags you want to train towards. What happens is the learning algorithm is supplied with the remaining tags + your embedding tag, and the associated image, it tries to generate an image based on that set of tags, then compares what the actual image is and what it generated to see what the difference is and learns from there, that difference is the loss. So short answer no don't add anything unless there's something in the images that your auto tagging isn't picking up and that you don't want to train towards.

If you're using automatic1111's web UI, the embeddings folder inside it, then you can use it in your prompts

It makes an obvious difference if it's being picked up or not, you could probably just prompt with the embed name by itself and either get something related or just a random girl or something depending what your model outputs with no prompt

I might have used a "my hands" hypernetwork on these, can't remember, but it'll be that if you've dropped it into PNG info to copy the prompt and if using the same model (which looks like it isn't on civitai anymore?)