Trigger word: @fhbpn
把画师 Boppin 的原创画风固化成单个 LoRA。干净的扁平动画风——清晰的黑色描边、平涂配简单渐变阴影、明亮通透的配色、大眼简化面部，背景常是白底或青海波一类的装饰纹样，人物体型偏娇小。这是我第一次尝试炼制画风 LoRA（也是第一次直接拿真人画师的作品当数据集，之前只炼过自己渲染的图），成色上还有进步空间，请多包涵。
底模需与训练底模一致（anima-base-v1.0）
提示词以
@fhbpn开场，权重 0.8 ~ 1.2（1.0 最接近原味，1.2 线条更明确、风格更浓）采样 30 步 / CFG 3.5~4.5 / er_sde + simple
务必接放大链路：一采之后用 SesquiLatentUpscale ×1.5 → 二采 10 步、denoise 0.35。这画风的看点在线条与材质层次上，小尺寸直出会把这些糊掉；放大之后观感差别很大
不要搭配加速类 LoRA（Turbo / Lightning 之类）。训练用的底模是纯底模，加速 LoRA 的效果没有参与训练，叠加只会让画面过锐偏色
About
A single LoRA that captures the illustration style of a popular Pixiv artist. The look is a clean, flat cel-shaded anime style — crisp black lineart, flat fills with simple gradient shading, bright and airy colors, large simplified eyes, and petite proportions. Backgrounds are usually plain white or carry a decorative pattern (seigaiha waves, woven texture, and the like).
This is my first attempt at training a style LoRA — and the first time I've used a human artist's work as the dataset rather than my own renders. Still rough in places; feedback is welcome.
How to use it
Base model: Anima base v1.0 — use the same base it was trained on, or the style will weaken
Start your prompt with
@fhbpn. Recommended weight 0.8 – 1.2 (1.0 is closest to the original art; 1.2 gives stronger, bolder linework)Sampling: 30 steps / CFG 3.5–4.5 / er_sde + simple
Use an upscale pass. The whole appeal of this style is in its linework and material detail, and a small direct generation flattens it out. Run SesquiLatentUpscale ×1.5, then a second pass at 10 steps with denoise 0.35 — the difference is substantial.
Do not stack accelerator LoRAs (Turbo, Lightning, etc.). Training ran on the plain base model, so their effect was never part of it; adding one on top only makes the image over-sharp and off-color.
Training details
265 public illustrations by the artist · resized to a 1536 long edge · WD14 auto-tagged with character names and style descriptors stripped out (so the trigger word owns the style) · dim 32 / alpha 16 / unet_lr 1e-4 / AdamW8bit / 3500 steps · final loss 0.0857.
Known limitations
Night and low-light scenes are noticeably weak — the dataset contains almost no night illustrations, so those prompts tend to collapse into silhouettes. Bold extreme-perspective compositions aren't reproduced reliably either. And the closer you want to get to the source's level of detail, the more you depend on the upscale pass above.