Part 3 of our 3 part Series of articles exploring LoRA Character Bases
In Part 3 we are going to put everything we have explored in the first two articles together into a seamless character creation flow inlcuding
- Designing
- Training
- Iterating
Follow along as we build our Space Paladin, Justine!
If you haven't worked through the first two parts of this series please take some time to familiarize yourself with them now:
Part 1: Base + Style
While making this series, we found a LOT of our LoRA's that do consistent character generation and work excellently when building a new character just with prompting!
Talk about easter eggs :) Take "Chibi + Fantasy Watercolors + Pastels" for example:
We got inspired when stumbling upon consistency with the model below.
Fun fact, these Composition Models were originally created as examples of how to blend LoRAs together. We were so happy with the results we released them as Platform Models.
We found a lot of hidden gems, particularly for Cartoon and Anime styles.
So we went ahead and generated images of a new and unique character using just the following prompt:
a celestial woman with pink hair and amber eyes in a white armor [+ description of mood, action or pose]
This created a really consistent dataset for us.
.
Next we uploaded the images and started a new training. We kept our captions short and sweet, and made sure to maintain consistency.
We also included a unique token and stuck to words that we might want to be able to prompt to guide the character again in the future.
More on training a character LoRA here:
Then we had our own Character Base LoRA!
But how to make her look more real?
First we did a quick prompting test, and found that it was useful to identify aspects such as hair color.
Next, we created a model using "Compose Model" and started with Justine as the base.
We tested her in a Chibi style:
It felt especially inspired when we tested a realistic character model, but we weren't exactly happy with all the details, such as the length of the hair. We also realized that we needed to prompt her pointed ears with realistic models to push them through:
We switched to our Naomi Character Base model, enjoying the soft background and her overall aesthetic, which we thought would pair nicely with some of the unique features of Justine.
We adjusted the strengths of the LoRAs, because we saw Justine was strongly pushing through, and opted to call her in with descriptive prompts rather than just pushing up her strength.
What do we mean by that? Because we used captions to associate her with things like "pink hair" and "elf", the AI model will naturally associate her with other features such as pointed ears, etc.
We also included the prompt "photoshoot style" and "dslr" to impose more photorealism, which is present in the Naomi model:
We still sensed that the model wasn't pushing through enough elements. We did not want to lose the quality of realism, so we went back and added the Space Witch LoRA in, to help impose the armor, but keep the softer details I associate with the character.
We often discover 3 LoRAs is the Sweet Spot.
And there you have it! That's how we can create an anime character and use Scenario to transform her into a realistic character, leveraging Composition and Character LoRAs!