What kind of prompts are you having difficulty with?

I'm not sure if an embedding can solve the problem. If there is any prompt combination that results in fixed necks, that can be turned into an embedding. I think the problem has to be finetuned out of the model, possibly with a negative finetune. however, finetuning isn't something I've been able to get good at yet, so hopefully someone else with an interest can look into it.

Can you guys give me example pictures of the long neck problem, and some example pictures where there is no long neck problem? From any model, doesn't matter. Pictures where it might not be obvious at a glance if the neck is too long or not.

You need to put the file you download into your embeddings folder, and then use the trigger word (the file name without the .pt) in the negative prompt.

The main post points already this out and informs people that the actual training images are under the version labeled "Actual Training Images".

I apologize for taking so long to reply to your comment- I had assumed that CivitAI would inform me any time a comment was made on one of my creations, but that may not be the case.

There are no captions for the images, as the caption used during training was from a template. I cannot recall what was in the template, but it was either "a picture of" or a list that was based off of the ones used in the original SD TI paper.

As for why I used 5000 images- I wanted the embedding to "average out" from the training images. If I put in too few images, then the embedding would take on the characteristics of a random sample and would be characteristic of that sample instead of the "true" characteristics it should have. An extreme example would be if you used only 1 or 2 images, the training program would have a difficult time learning what you are trying to teach. The more images you present, the more certain you can be that the final product is representative of the whole rather than of the few. This thinking does have a limitation- I do not actually know what is too few or too much. I chose 5000 images because early TI tutorials would suggest training an embedding to 5000 steps. I figured that if I had a unique image for every step, it would provide an optimal outcome. It could be that the ideal is 500 images or 50000 or some other number. Methodically testing such things is difficult due to the amount of subjectivity involved.

The reason why using significantly less images is a good idea when training a normal embedding, and other things such as LoRAs, I think is due to the amount of extraneous concepts that end up getting included the more pictures you add. The more concepts you introduce, the more "good" pictures you need to include in order to properly teach that concept as well so that it can be adequately differentiated from the other concepts. What is also majorly problematic is that we (people) tend to not realize how many new concepts we might be introducing with the images that we are using, and don't tag the images accordingly or discard problematic images. This information is severely underrepresented in tutorials- the furthest the author will usually go is to say what I have said already- you need "good" images and "good" tagging.

Anyways, I trained the embedding the way I did as a way to sidestep needing to both find and tag images. Sometimes it works, sometimes it doesn't.

I trained the negative embedding using the actual training images. They are there in case someone wants to use them to train a negative embedding using their own settings.