<h2>底模</h2>anima<h2 id="trigger-word">触发词（Trigger Word）</h2><pre><code>tongyao</code></pre><h2 id="overview">Overview</h2>Tongyao is an original character trained as a character LoRA on the Anima base model.She is a gentle-looking girl with long silver hair in twin braids, blue eyes, and small blue rosestucked into her braids. Her default outfit is a white nun-style dress with gold trim and a goldcross pendant, worn over white thighhighs.This LoRA is designed to keep her face, hairstyle and ornaments consistent across very differentscenes — bright outdoor backgrounds, low-light interiors, dramatic backlight, and so on.## Trigger Word<code>tongyao</code>Put it at the beginning of the positive prompt. That alone is usually enough to pull in her faceand hair. Clothing, pose and background should be described separately, because this LoRA wastrained mainly on the character herself, not on a fixed outfit.## Recommended Settings| Setting | Value ||---|---|| Base model | <code>anima-base-v1.0</code> || LoRA weight | 0.8 – 1.0 || Steps | 30 || CFG | 4.5 || Sampler | <code>er_sde</code> || Scheduler | <code>simple</code> || Resolution (portrait) | 896 × 1152 || Resolution (landscape) | 1216 × 832 |## Prompting Tips- Tags first, prose second. Tags at the front, then a short natural-language paragraph for light, camera angle and mood.- Camera angle needs prose. A bare <code>from below</code> tag is often ignored. Writing "the camera sits almost at the water surface, tilted steeply upward" actually moves it.- Light needs prose too. <code>backlighting</code> alone is weak; describe where the light comes from and what it hits.- Outfit is prompt-driven. The LoRA holds her identity, not her clothes.- Suggested negative: <code>lipstick, makeup, realistic, 3d, flat lighting</code>## Training| Item | Value ||---|---|| Output name | <code>anima_character_v4</code> || Steps | 2500 || Epochs | 70 || Rank | 32 || Base | Anima |## Notes- Original character. Please credit if you repost.- Trained for Anima; results on other base models will differ noticeably.- Trained mostly without fixed backgrounds, so simple / plain backgrounds are the most stable.

Files with SHA256: 9da5795292d3a9d2267c9d978f0eda83474044f8e7290d48f121f8ccb4eb8d5b