Sign In
homemodelsimagesvideos3D Modelsarticlescomicschallengesupdatesshop

Tagging for Creatures Tested: “pokemon (creature)” vs. “creature” vs. null

7

NanashiAnon's Avatar

NanashiAnon

Silver Writer Badge

Jul 27, 2025

(Updated: 4 months ago)

training guide
tag comparisontagging
Tagging for Creatures Tested: “pokemon (creature)” vs. “creature” vs. null

After making my model for Golden Sun's Venus Djinn I realized I should figure out what works best before I try the other 3. Having buzz to spend and a dataset I submitted the same dataset for training, with identical settings, changing only the tags. The first (original) used “pokemon (creature)” and “pokemon on shoulder” in its two instances in dataset. The second just removed “pokemon (creature)” with no replacement (but left pokemon on shoulder because I forgot) . The third used “creature” and “creature on shoulder”.

Observations:

  1. Spine count just isn’t going to be correct consistently. Exact counts of something is something SD models are bad at in general, no matter how iconic and consistent the number of buttons or rings in vambraces. Likely related to why it struggles with humans having two arms with 5 fingers.
    Some identifier of non-human stats is needed. Version two did really poorly with general shape

  2. All took 5/6 epoches to be recognizable. This is generally more than humans need with same training settings.

  3. “Pokemon (creature)” really likes to make things into Pikachu, especially other Pokemon tags like Pokemon on Shoulder. A negative can help here, but it’s obvious here.

  4. “pokemon (creature)” handles being the subject of the prompt better than just creature. One was consistently putting putting the djinn on the chair, the other was consistently imagining a humanized thing sitting on the chair with the djinn in the background. (test finished model with no humans)

  5. “creature” makes the subject look rather fat, and I can’t seem to control it with negatives

I conclude that "pokemon (creature)" is likely the best general purpose

Doing this comparison blew a good chunk of buzz (1000 extra). I’ve got a stockpile and I don’t want tips, but I would love it if you went to my LoRAs, used them and shared them with everyone using the post button in the gallery (not post in generator or new post under create). This gets me 50 buzz (1/10th the cost of training a model each attempt) per model you post to this way (but only once each model, I’m not sure if posts to different versions of a model gets me buzz)

7