This LoRA is intended to mimic the style of the artists BHM / Trashup / Muchu2Muchu (all the same artist, just under different names). His artstyle has evolved noticeable over the years, so the intent is to have two alternate versions, one to mimic his old style and another to mimic his the earlier version of his current style, which I'll call "intermediate".

The break point I'm setting between "old" and "new" style is how glossy the style is. Not long after getting banned from deviantart (that's where the Trashup username is from) his style changed and became more detailed and "glossier" for lack of a better word.

Example of "old" style

Example of "intermediate" style

Example of current style

I don't plan to do a version on his current style (which has dialed down the glossiness but kept the detail), that feels like a rather mean thing to do to a relatively obscure artist without getting permission.

intermediate-v0.2-Pony & intermediate-v0.2-IL (IL is Illustrious)

Tsubaki_BHM is maid, Emily_BHM is nurse (found her name in some old art)

These versions have more training data to reduce over-training on specific images and increased epochs to better learn the characters and concepts.

See the trigger words for what concepts these covers beyond the style. The Illustrious version can attempt tentacle coils with "coil," but it's not great at them.

intermediate-v0.1.1-IL - now for Illustrious

This is the v0.1 training data but with increase number of epochs to better learn the characters & concepts. This worked, but also resulted in overtraining.

intermediate-style_v0.1 - my first pass at BHM's style from about 2015 until roughly 2019 or 2020. Trigger word for the style is still "BHM", Maid's trigger remains "Tsubaki_BHM", but now I've added "Nurse_BHM" and "slug monster" which sort of work, though additional details help with Nurse_BHM (see trigger words) and "slug_monster" tends to produce a yellow blob with tentacles. Suppressing low quality tags helps. Strengths of 0.7 - 1.0 seem to work well, depending on the number of other LoRAs being used.

I've found that stacking both versions with strengths of 0.75 (intermediate) and 0.5 (old) can give some interesting results, see the example with Rabiane.

old-style_v0.1 - my first pass at this, didn't come out terribly well (I think), but more testing is required. Trigger word is "BHM" for the style and "Tsubaki_BHM" for his maid OC (doesn't get her hair color right, unfortunately). Sub-style / concept trigger are "mucus" and "tentacles", though you likely want to supplement with additional LoRAs to help with each specific concept.

This version seems to still be decently effective at strengths as low as 0.5, which seems to degrade the image quality less than at 1.0.

I didn't do a great job with the score tags in training, so using "score_9, score_8_up, score_7_up" in the positives and "score_6, score_5, score_4" in the negatives helps improve the results.