Why does a trained model on a person start generating other people if I specify "full body"

[juan9999]

I've trained a fast-DreamBooth colab with 10 images of a person, mostly faces, 2 half body, 1 full body. 500 training steps, 5e-6, 250 encoder steps, 1e-6 Text_Encoder_Learning_Rate, using a 1.5 model.

I then generated some prompts using:
portrait of tall skinny ((RLTYRTTRN)) in tulum, boho style, lots of boho jewelry

I get pretty decent image generation, with potentially an overtrained model since I seem to get very similar body positions as the training set.

If I change the prompt to
full body portrait of tall skinny ((RLTYRTTRN)) in tulum, boho style, lots of boho jewelry

I then get images generated of another person/ face sometimes even another race!

The closeup is the face that I trained, the one with blank pants is not even close!

What am I doing wrong and what do I need to do to get the same face/body for full images as I do for close ups?