In today's experiment we are going to make a Lora for an OC Pokemon with only 8 images!!
Disclaimer: this is not a guide its more like my research notes. I was able to create a good Lora with this example but it may not be so easy depending on the complexity of another project
Credits to the original arts:
OC design by Onemegawatt - https://twitter.com/onemegawatt
3D model by Pauline De Houwer - https://skfb.ly/onuqA
Steps
Find/Create a 3D model for your OC
Collect a few images that can can quickly be edited to look like your OC
Screenshot the 3D model for a few different angles: Front, side, back, close use, far away, etc
Collect any already existing art of this OC if they exist and/or use AI to generate some pictures
Make simply edits to existing art to make it look like your OC (it doesn't need to look perfect the AI just needs an idea of what to make)
Train Lora with Dataset!
Step 3 - Screenshot 3D model
Here are the screenshots I took of the 3D model https://skfb.ly/onuqA
Full resolution picture: https://imgur.com/Xe6NMTu
Step 4 - Already existing art and AI generated art
On the left, you can see the OC art that started it all
On the right, you can see I did a simple AI image to image to clean up the lines of one of the 3D Model screenshots.
Full resolution picture: https://imgur.com/uLaObZd
Link to the art by Onemegawatt: https://www.pinterest.com/pin/175710822959409369/ *
*unfortunate;y I can't find the original post so here is a link to the artist's twitter: https://twitter.com/onemegawatt
Step 5 - Simple Edits
Here you can see I simply used bucket fill to recolor the images (The pink color palettes is Skitty's official colors).
I did draw in the fangs and purple spikes on Skitty's tail but as you can see its really sloppy.
On the left is the existing art, on the right are my simply edits
Full resolution picture: https://imgur.com/m9mY1qE
Step 6 - Training and Conclusions
Most of my Pokemon/Create Loras are trained with these settings
Epochs: 10
Number of Repeats: 10
Network Dim: 32
Alpha: 16 (Alpha is usually half of the Dim)
Flip Aug: true
As you can see my Shadow Skitty looks more fluffy than the original Skitty: this is by design but its possibly due to the fact that the outlines in the edits are not very precise. Like I said in step 5 I was pretty sloppy with my edits and the AI may have interpreted this as "fuzzy hair"
I've said this before in other experiment notes but I'll reiterate it: the key to success on this was the 3D model.
3D models are essential for getting lots of good pictures and angles.
Closing Comments
If you liked this article please give it a like and feedback. I'll continue to document my research if people show they are interested!
Also please consider liking and posting pictures of my Shadow Skitty Lora: https://civitai.com/models/161770