Style LoRA trained on NoobAI-XL V-Pred 1.0
I tried to make an anime LoRA that doesn't look exactly like an anime screencap. It looks a little less flat and has a 2.5d feeling to it.
Recommendations:
Cfg/Guidance: 3 (softer) - 5 (sharper)
Steps: 15+
Strength: 0.7 - 0.8
Resolution: 1024x1536 (vertical) - 1536x1024 (horizontal)
Update 1.2
Changes:
New dataset (Mostly NSFW)
More stable
Neutral lighting (can be changed with prompts)
Less artifacts
Better prompt adherence
Less saturated with higher cfg/guidance
Still not perfect with hands (its probably the 1024x1536 resolution, use 832x1216 when needed)
I recently discovered that DMD2 LoRAs can work great even with only 4 steps and can generate good quality images for certain animes.
Update 1.1
Some changes should be:
Larger dataset
Better prompt adherence
Less yellow lighting
I noticed this version generates bad hands very often, I am working on a new version to fix this.
Trained on both SFW and NSFW images