This is a concept model that adds colored skin training. I noticed a common problem for untrained lora/checkpoints that it tends to make patchwork/unmatching skin without proper training. Tends to work best at 0.6. This model was trained on NAI anime-full-pruned and generated in Anylora. Made with the Kohya tool for AI Training through google colab. Any suggestions to improve or recommended things for the data set would be appreciated.
Description
This was trained on over 150 images but some colors involved in training still need to have more samples for study. The weakest trained color in the mix is yellow it also doesn't help that it's close to a normal skin tone. Brown (as in art pallet brown not traditional skin tone) has minimal samples but seems to have been understood well. Pink and purple are dull and need to be revamped with darker colored characters to make a distinct difference. Any suggested characters or content for any of these colors is appreciated.
All colors included are:
-black
-green
-blue
-white
-grey
-green
-red
-brown*
-purple*
-pink*
-yellow*
*weak
FAQ
Comments (4)
May I ask how did you pick images so the training model focused on skin color? seems a difficult thing for the AI to pick on
How did I pick images? Colors were separated into folders to split up the training each folder had the goal of at least 20 images. The colors that i couldn't find enough of are weaker to the others like yellow I believe has the least and as a result isn't very strong when prompted if i remember correctly it had like 8-12. The images for training had to be obviously the designated color so no sketches or low res images were used. Just detailed colored characters. To improve I'm going to have to expand the training images and cut back on repeated characters. Having the same character too many times makes it so when there skin type is prompted the picture will slightly look like them which Is not the goal.
Was working good, but seems to not like the EasyNegative embedding.
It really struggles to do males.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.