This style LoRA was trained using pure illustration source images, excluding any anime materials.
For best results, use a weight of 0.7–1.0 when using this LoRA alone. When combining it with other style LoRAs, feel free to lower the weight as needed. This LoRA is great for enhancing the realism and texture of the skin.
Prompts
artist:betmillustrationstyle,Please format your prompts using trigger words followed by natural language. All example images include ComfyUI workflows embedded and can be dragged directly into ComfyUI for use.
V2.0➡️Adjusted the training parameters again and trained for more steps; the art style and details are now much more faithful to the original.
V1.0 ➡️Streamlined the dataset from 210+ to 110+, moderately increased character diversity in the images, fixed the issue where every generation produced pink-haired Yuji Itadori regardless of the prompt, and completed a full training run using tested steps and parameters.
V1.0test ➡️This is a test version primarily aimed at testing MOSS's ability to replicate the art style. I'm very satisfied with the results. However, since the dataset tagging inherits Anima's scheme, the model's text comprehension might be slightly reduced.