Thank you so much for this, and looking forward to seeing how it improves over time. I use SD to make tokens a lot for my game, and when I tried to train a LoRA myself it turned out awful. This really helps get that perspective much easier.

So nice and helpful. Very cool helps to visually diversify the battle scene.
I will definitely show my results when I generate tokens for a batch of my players. I got excellent bandit tokens with daggers. But, I can’t generate a paladin with a two-handed sword, the model was not trained on a two-handed grip? I tried greatsword, great sword, two-handed sword, long sword, longsword, claymore, zweihander, bastard sword, espadon, two handed grip, two handed sword stance.
Also, I have problems cutting out tokens from a black background, for some reason it turns out uneven, and “selecting an object” in Photoshop can’t be done, I have to do it manually (and I’m already so unaccustomed to this). So far I haven't been able to automate this. The variety of tokens is amazing, but cutting them out is still a bit tedious.
I also had problems with the scimitar, halberd, and crossbow,. The model sometimes generates curved blades without a request, and with a scimitar request, it generates both curved and straight swords.
Once, the model generated a token on a white background. I added a “white background” to the prompt, but it didn’t appear much more often.
I also tried generating lizardfolk and tabaxi. Lizardfolk often came out looking like goblins, bugbears, or abstract furry furries, and had long hair. Tabaxi turned out much better, although sometimes they looked like tanuki.
Maybe it’s worth trying to train the model not on tokens with a transparent (black background) but on a white background?
This is a very useful tool, I hope that you will develop it further. If I can help you in any way, please contact me.

Very nice work, it requires a bit of mastery and finesse but you can get very nice results.

If you plan to do it on flux it would be incredible. I made an img to img from this lora sd xl to flux and the result is really outstanding.

So, all the figures have black lineart, right? But you also trained it to make the background black as well. So now there's no way to separate the background from the line art.