A Illustrious LoRa for Shuvi Dola from NGNL Zero, because it did not exist in a capacity I was happy with.
After several all-nighters, V2 is done! That's pretty awesome, if you ask me.
Recommended weights:
Between 0.8 and 1Recommended tags:
Shuvi, cable tails, brown hoodie, mechanical wings, mechanical partsTrained on Kohya_ss & Civit AI
V2 Changes:
Massive part quality improvement
Hips now generate correctly
Metal cheek plates more consistent!
Wings included in data set (trigger: mechanical wings)
Iconic brown hoodie fixed (trigger: brown hoodie, hood up)
Cable tails added (trigger: cable tails)
The hair ornament can be forced to generate with the prompt "hair ornament"
No longer generates human ears
Known issues:
The mechanical array running down the left side of the torso will almost never generate while in revealing clothing.
Mechanical parts may generate over too much space
Removing clothing may result in "biologically accurate" features (rarely)
Mechanical parts may overlap clothing when generating (rarely)
Description
FAQ
Comments (6)
LoRa roadmap:
17 base images from the original training set (to be filtered into the following categories:)
Side angles, with emphasis on the right side detail (10 for each side)
Back angles (15)
10 close-ups of the face
Various poses (15)
Various outfits (10-20)*
I’d like some of the hips specifically (seeing as the model tends to forget the visible mechanics), but we’ll see (5)
=75-85 total
*10 to emphasize the positioning of the mechanical parts true positions, another 10 for the default outfit.
Hoping it'll be done by February 15th, but midterm season being the way it is: It may be delayed.
Additionally, if you get some particularly banging results, DM me! I'll gladly include it in the training data :)
V2 is showing promising results thus far, which is genuinely surprising- Given how there's no immediately obvious way to make Kohya_ss train for illustrious yet. All things considered, I'm happy with the progress made today.
Progress report- Day 3
Insert head_hitting_desk.sfx here.
Tried making a LoRa locally, and it actually works really well... On SDXL.
This is an illustrious model.
My illustrious checkpoint completely ignores the data.
...Tomorrow is a new day. Or rather, today seeing as it's 3 am.
Say, anybody want to buy me a coffee? Lol.
V2 is functional on illustrious checkpoints.
Progress report- Day 4
So, the local LoRA creation process is at least functional now, and I could release V2 today- If I really wanted.
But, here's the thing: Do I release a potentially overtrained LoRA that I'm not entirely happy with?
...Of course not.
I didn't think this would be going so fast as-is, so that just doubles the reasons why I shouldn't push it out early. There's no Kohya_SS guides for illustrious training, and I'm definitely feeling the monte carlo method now.
Ah, but with all things, patience is required, so it'll probably be a few days yet.