这是一个风格lora，灵感来自 @purplelady 的 Arcane Violet，我真的很喜欢那个 lora，可惜只有 Anima 版本，我自己收集了素材训练了这个 Lora，注意，该 Lora 基于 Mose Base 模型训练，在 Base 上的表现会比 SFT 好很多，如果你不喜欢跑50步，推荐使用 https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd。
This is a style LoRA for MOSS-Image.
It was inspired by @purplelady 's Arcane Violet, a LoRA I really loved — unfortunately it was only available as an Anima version. So I collected my own reference set and trained this LoRA from scratch.
Train on Base, use on Base. This LoRA was trained on the MOSS-Image Base checkpoint. It performs noticeably better on Base than on SFT, since SFT's aesthetic post-tuning dilutes the LoRA's style signal. Use it with the Base model, not SFT.
Fast sampling (recommended) If you don't want to run 50 steps, pair it with: https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd That turbo LoRA is distilled from the Base model, so it stacks cleanly with this style LoRA — about 6 s/image at 8 steps with no quality loss.
Recommended sampling parameters (settled after comparison, beats the initial set): euler + simple/sgm_uniform on 50 steps or 8 steps with 8steps lora.
No trigger word. It is an unconditional style LoRA — no trigger needed, just adjust weight.