Model was trained on D&D monsters, heroes and NPCs. So it works best for medieval fantasy setting but not limited.

!IMPORTANT!

Usually for FLUX is recommended to keep CFG (Not Distilled CFG) 1. But for this Lora I recommend to increase it to 2, but you can play around and try different values. Also in this case negative prompt works.

Dimensions

Model was trained on 768x768 resolution, however 1024x1024 should works pretty well. If you trying something not square like some beast you can try 768x1024:

("Topdown view on a fire dire wolf elemetal beast")

Prompting

In most cases i recommend natural language prompting instead of tag-based prompting. For example instead of "Humanoid, orc, greatsword, hide armor" Use "humanoid orc male holding greatsword and wearing hide armor".
Also in many cases it is better to start prompt with "Topdown view on a...", for example "Topdown view on a humanoid orc male holding greatsword and wearing hide armor". Both of these tips makes result more detailed but sometimes it can make it too far from training dataset, so you can try different prompting depending on subject.

"Topdown view on a humanoid orc male holding greatsword and wearing hide armor":

****By the way, if you are my subscriber you can notice that I used specific weapon and armor name and it worked (more or less). I used these words as tags while training, so go try it.

Still two-handed weapons are not generating stable. But you can use tag "Dual" for two weapons in different hands. For example "Dual daggers" or "Dual shortswords"

Same as for previous models, you can try creature type like "Undead", "Fiend", "Beast", etc but they works not so good at this model.

For some reason I did not manage to make model learn new concepts and it is recommended to describe additional details. For example instead of "Tiefling..." use "Tiefling with red skin, pair of horns and tail"

I did not expect that, but you can even...

Other settings

I recommend to keep strength close to 1.0.

If you can wait, 30 steps is significantly better than 20 steps.

You can also try increase Distilled CFG Scale little bit (I prefer 4.0) but it is not so important.