docicsxl
Style mix LoRA based on Dan Mora + Celestin.
Trigger word: docicsxl
This is a personal style mix aiming to blend Dan Mora’s linework with Celestin’s cleaner rendering, while reducing the heavy orange color grading that Mora’s colorists often use.
I experimented with tagging things like Ranger Slayer, Superhero Costume, and Super Sentai mostly for fun — Ranger Slayer didn’t fully learn, but the superhero tags ended up creating some interesting mash-up costumes.
The LoRA tends to:
Slightly change style when using superhero costume or Power Rangers tags
Add chest logos when using bodysuit + Power Rangers tags
Stay stable up to high CFG (surprisingly doesn’t melt after 7)
If characters start appearing in the background, add these to negative:
multiple girls
multiple boys
speech bubble
english text
comic
Recommended Settings
Weight: 1.0 (Works from 0.35 – 2.0, I mostly use 1.0)
CFG: 2 – 10 Sweet spot for me: 6
Feel free to experiment with both weight and CFG.
It’s not perfect, but I’m happy with where it landed after 7 attempts.
Training Details (for the nerds)
Images: 1184
Repeats: different per set
Epochs: 14
Steps: 3806
Batch size: 7
UNet LR: 0.0004
Text Encoder LR: 0.0001
Dim / Alpha: 64 / 32
Optimizer: Adafactor
Noise offset: 0.05
PyTorch memory enabling + Kyoha batch handling was a pain in the ass, but it worked.
