v1.5:

Same dataset, same captions, but different config and tool.

Used Citron Legacy's LoRA training ui (https://github.com/citronlegacy/citron-anima-lora-trainer-ui), config is in additional files for this version near the dataset.
Most notisable differences - 6400 steps instead of 2000 on CivitAI (I'm not shure, thay may count them differently, but in both cases it seems to be total amount of images*steps*epochs - 64*10*10) and scheduler functions.

As a result - better flexibility. Finally it is not fighting over clothed belly or different poses. Still not perfect, but already better.

v1.0:

Dataset of 64 images, mostly synthetic. Standart CivitAI trainer. Captions done in google colab notebook with 'SmilingWolf/wd-eva02-large-tagger-v3' and 'huihui-ai/Qwen3-VL-4B-Instruct-abliterated' (see notebook here). Result is good, but part with orbs sinking in is lacking for now.

Trigger word should always be first or after new line (qwen tokenizer splits it differently at start and in the middle of text) - '^soul pregnancy, etc, etc,' or '\nsoul pregnancy, etc, etc'.

Be careful with prompts - generic '1girl' would often generate a girl, not an adult woman. Add more tags for age or use other way to describe character.

Base: Another nieche concept, so nieche, that danbooru does not even has separate tag for it and r34 only has ~10 images.
Main idea is to depict a character with pregnant belly, that is invaded by spirit orbs (ghosts, spirits, hitodama, will-o'-wisps etc.).
One similar lora does already exist for Illustrious by @Aswoll, if anyone is interrested, but it have bias towards anthro/furry.