3 months late but for anyone else what has helped me to guide models:
1) First thing's first, use prompt plotting (try with and without this lora with the same prompt) to rule out if this lora has a dominating preference for a character in a particular position. MANY loras have this problem where it forces a dominant position, especially at high strengths. If you get the right position with no lora, that is the culprit. Try incrementally increasing the lora strength until it starts forcing the wrong position and then you know your rough limit.
2) Picture the image you want in your mind and focus on what you can SEE only in that image. If you can't see a certain body part, certain characteristics, then DON'T prompt for it. For example, if you want to see a character from behind, do not have "collarbone" in the prompt because you'd only see that from the front and the model will be guided by that. Instead use something like "shoulder blades".
3) Start with the best possible prompt guidance (you often need multiple queues to the model to reinforce what you want), check results, see if you can add more guidance and/or start boosting the most critical criteria of the prompt with weights.
4) Don't ignore negative prompting but it should be a last resort after you have enough positive prompt guidance.

In your case, I have not had problems forcing that position or even needing to force that position. In fact usually my issue is having a hard time getting the character on their back. Therefore I have no quick and easy suggestions that I know will work.

Personally I'd try adding "on back", "lying on back", "looking up" or "lying on back looking up" to the negative prompt. Sometimes the added context helps over super short contexts like just "looking up" because it will simply turn the head to the side instead of understanding the broader context.

Before even adding the negatives though, I'd try "lying on stomach" or variations of that. Then "lying on breasts", "breast rest", breast pillow".

Hopefully other people see this because this knowledge applies to any lora and most models until prompt adherence improves significantly. (Z-Image?)

If you're still struggling with this, try "breast press" or "breasts on floor".