Anima ver

Satisfied result after 2 training attempts. First time training lora on civitai. Used 72 images, prioritized style consistency over pose variety, default 1500 training steps and parameters. 1st attempt at 1900 steps only has 1 out of 10 epochs is good but color too muted and no ^^^. 2nd attempt at default steps and what you see is the best out of 10 epochs.

Turns out he is using novelai instead. His recipe: Hews, Cutesexyrobutts, Shexyo, Rucarachi, pale colors, etc. Current similar style on pixiv Baker Arts

Suggestion: Add "shadow" to enhance the style by highlighting the contrast between shaded areas and shadow for another variant of the art style.

Planning to retrain the lora using 1440 steps, 10 epochs, 0.0001, cosine, 16, 8. Plus, I only found out using Base v1.0 comes out so bad that somehow me always using turbo was unaware of this lora's flaw. But I didn't download the dataset with tags (which took me about 3 days to tag everything) used in the first training, so it's tagging hell all over again.

Illustrious ver
Trained using 100 images from Truck kun 2025 June image set.

This lora can still imitate the style to most degree but not very close due to the dataset using images that have slight different styles here and there among other different styles (which I didn't notice when I was newbie).

Weight: 1.0-1.3. Either Euler or Euler a is fine, but preferably Euler checkpoints.

Style depends on CFG scale but depends on checkpoints
- high for that exact heavy outlines style
- low for pretty much subtle enough for other style.

See cover images for references. Generated using VXP_ASC vPred (Anime ScreenCap) [Only on PixAi] but you see that can this lora can replicate the style with the right checkpoints.