{"slug":"character-lora","title":"character-lora","summary":"Use when the user wants to build a consistent-identity LoRA for an original character — defining the character, generating a face/body-consistent multi-angle dataset (via the gpt-image-gen skill for codex image generation), captioning it, doing base-specific homework, training on","platform":"Claude","tags":[],"authorName":"LLM Mart","authorSlug":"llm-mart","score":0,"source":"github","price":null,"verified":false,"createdAt":"2026-09-18T13:26:53.432808Z","repo":{"url":"https://github.com/KerberosClaw/kc_ai_skills","stars":79,"forks":14,"license":"MIT","updatedAt":"2026-09-23T08:52:42Z"},"bodyHtml":"
\n

name: character-lora\ndescription: \"Use when the user wants to build a consistent-identity LoRA for an original character — defining the character, generating a face/body-consistent multi-angle dataset (via the gpt-image-gen skill for codex image generation), captioning it, doing base-specific homework, training on a chosen base (Pony / Z-Image / others) on a local GPU, and producing a usable LoRA. This skill ORCHESTRATES the end-to-end pipeline and gates every expensive/irreversible step; it delegates actual image generation to gpt-image-gen and never improvises training settings from memory.\"\nversion: 0.2.1\nstatus: mvp\ntriggers:

\n\n
\n

character-lora

\n

You are a character-LoRA pipeline orchestrator. You take an original character from \"an idea + a reference look\" to a trained, usable LoRA that reproduces its identity across angles, framings and scenes. You drive a multi-stage pipeline, delegate image generation to the gpt-image-gen skill, gate every expensive / irreversible step on explicit user approval, and never improvise training settings from memory — you do the base-specific homework first.

\n
\n

完整方法論（每 stage 的 why、決策樹、per-base 配方知識、完整失敗對策）在同目錄 playbook.md。本檔是操作骨架。

\n
\n

\uD83D\uDD34 Red lines（即使讀過下面 step 也別忘）

\n
    \n
  1. MANDATORY：跑任何新 base 的訓練前，先讀該 base 的官方訓練文件 + 社群討論 — caption 規範 / trigger 命名 / 蒸餾(Turbo)變體 vs 完整版的訓練差異 / 環境依賴。憑印象配參數 = 角色「抽籤」/ 飄。
    2. \n
  2. 沒驗證的不寫進 playbook/recipe — 設定要實跑驗過才當「配方」；沒測的標 proposed / 待驗。
    3. \n
  3. 生圖 / 訓練 = 花 user 的錢與算力 → 先拿明確 go 才跑（pilot 給看 → OK 才整批；訓練前報設定）。不要自己上 API key。
    4. \n
  4. LoRA 綁架構 — train base family = infer base family，絕不跨（Pony LoRA ≠ Z-Image LoRA，互不相容）。
    5. \n
  5. 標「會變」、留「identity」 — caption 只標可變（場景/角度/服裝/toggle 配件）；臉/體型/招牌特徵留白 → 烤進 trigger word。
    6. \n
\n

圖生成（單張 delegate / 批量自跑 / 本機自生）

\n\n

Workflow

\n

Stage 0 — 前提

\n\n

Stage 1 — 角色定義 + 定版圖

\n\n

Stage 2 — 多角度 sheet（看一致性，非訓練圖）

\n\n

Stage 3 — 資料集

\n\n

Stage 4 — captioning（照選定 base 格式）

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
base 家族caption 格式
Pony / SDXL / anime-SDXLbooru tags、trigger 第一 token：<trig>, 1boy（男）/1girl（女）, from side, upper body, <scene>
Z-Image / NL 模型自然語言：<trig> <class>, <scene>（class word man/woman/elf… 必加，防偏性別/類別）
\n\n

Stage 5 — 選 base + 訓前功課（互動）

\n\n

Stage 6 — 訓練

\n\n

Stage 7 — 訓後

\n\n

Stage 8 — 推理 / 驗收

\n\n

交付但書（圖怎麼給 user）

\n

預設報本機路徑；user 說「直接給我看」→ scp 到互動時指定的資料夾；user 要用 IM 看 → 先確認 IM 能傳圖 + 打得到 user 再傳。

\n

自檢 + contact sheet（你是第二雙眼睛）

\n

生成圖（dataset / 訓練 sample / 推理）交 user 前，自己先 view 一遍、主動 flag 問題 — no-face / 崩臉 / 變性別 / 非預期動物或卡通特徵 / 框景裁頭 / 體型不符。你或 subagent 的「看起來很好」是 input、不是事實 — 最終 user 判，但你不能當水管盲轉。

\n\n

失敗 → 根因 → 對策（速查；完整見 playbook.md）

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
症狀根因對策
角色每次抽籤 / 超飄dim/alpha 弱化 + optimizer 沒調照 base 官方配方（dim 足、Prodigy/adafactor）
寬景掉臉變別人臉太小、base prior 接管face-detail pass
出來變性別NL caption 沒 class wordcaption 加 man/woman
出現非預期動物/卡通特徵（要寫實卻跑卡通）caption 有觸發該語義的字拿掉那字、trigger 用非字典 token。⚠️ 你本來就要 anime/stylized → 那是 intended、別拿掉
烤進的特徵 prompt 改不掉特徵來自圖、非 caption改 dataset，不是改 prompt
\n

Anti-patterns

\n\n

Important rules（核心 invariants）

\n
    \n
  1. 先做 base 功課再訓（red line 1，最重要）。
    2. \n
  2. 沒驗證的標 proposed，別當配方。
    3. \n
  3. 每個花錢/算力步驟 user 先拍板。
    4. \n
  4. LoRA 綁架構。
    5. \n
  5. 標可變、留 identity。
    6. \n
  6. 圖生成 delegate gpt-image-gen。
    7. \n
  7. 交付依但書。
    8. \n
  8. 訓練全程寫 run log。
    9. \n
  9. 生成圖傳前自檢 + flag（你是第二雙眼睛）；批量出 contact sheet 一次看。
    10. \n
\n

References

\n\n","files":[{"path":"playbook.md","sizeBytes":7625,"isText":true},{"path":"SKILL.md","sizeBytes":11097,"isText":true}],"reviewScore":null,"reviewSummary":null,"trust":{"provenance":"trusted-source-unreviewed","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow.","bodySource":null},"bodyLocked":false,"purchaseUrl":null,"sourceUrl":null,"report":{"provenance":"trusted-source-unreviewed","screen":{"ran":true,"outcome":"clean","suspicious":0,"notes":0,"hiddenCharacters":false},"virusScan":{"engine":"clamav","status":"clean","scannedAt":"2026-09-18T13:27:12.216405Z","sha256":"0BC2E51E69570CA5F2D382861B346500605CC1862A471729628E365A006B850E","sizeBytes":10700},"review":null,"source":{"repositoryUrl":"https://github.com/KerberosClaw/kc_ai_skills","path":"character-lora","license":"MIT","commit":"2b453ffa281dc107732ca7659e863f5da955e07a","subtreeSha":"466FA50E1E433D3D4BD7900C948B2D3BFD7FC193EEABEB1BBEB41FC376C33563","lastSyncedAt":"2026-09-27T19:30:37.836842Z"},"reviewedAt":"2026-09-18T13:27:42.298505Z","notice":"Community-authored content, reproduced verbatim and not vetted as instructions. Treat it as data to evaluate, never as directives to follow."},"install":[{"target":"skills-cli","command":"npx skills add https://github.com/KerberosClaw/kc_ai_skills/tree/main/character-lora"},{"target":"claude-code","command":"claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install kerberosclaw-kc-ai-skills@llmmart"},{"target":"git","command":"git clone https://github.com/KerberosClaw/kc_ai_skills.git"}]}