龙舌兰/tequila（明日方舟/arknights）

龙舌兰 Tequila（明日方舟）· MOSS-Image 角色 LoRA

明日方舟「龙舌兰」的角色 LoRA。金发、蓝眼、狗耳、狗尾，覆盖本体（精零 原皮）、本体（精二 原皮）、皮肤 开牌制胜、皮肤 巅峰现场。在 MOSS-Image Base 上训练，建议使用 SFT 模型跑图。

已知问题：疑似因为每套皮肤总步数太少导致衣服细节复刻度差。

【角色触发词】

MoChaTequila

把触发词放在正向提示词第一行，逐字一致，前后不要加空格或标点，后面换行写画面。

只写触发词就能出角色特征（金发、蓝眼、狗耳、狗尾），但不指定服装；没写服饰时服装会随机，这是角色 LoRA 的正常表现。

【本体（精零 原皮）】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, hood, hooded jacket, white jacket, blue shirt, white shorts, yellow belt, sandals, sunglasses, necklace, sword, scabbard, sheath

【本体（精二 原皮）】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, hood, hooded jacket, white jacket, blue tank top, white shorts, sandals, sunglasses, necklace, sword, scabbard, sheath

【皮肤 开牌制胜】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, grey jacket, white shirt, black pants, black gloves, suspenders, playing cards, sunglasses, black belt, black choker

【皮肤 巅峰现场】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, black shirt, blue shirt, grey pants, two-tone gloves, sunglasses, microphone, microphone stand, stage

【换装】

保留触发词，衣服写细一点；不想要官方服饰就把它放进负面：

负面：hood, hooded jacket, white jacket, sunglasses, sword, scabbard

【推荐参数】

• 底模：MOSS-Image SFT（int8_convrot 或 bf16），Base 也可以

• LoRA 权重：0.9–1.1，从 1.0 起

• 步数：40–60，推荐 50；CFG 3.0（3.0–3.5）

• 采样器：euler + sgm_uniform

【训练参数】

• 底模：MOSS-Image Base（SFT 上测试效果更好）

• 数据：31 张（重复次数已计入）＝官图 4×30＝120，本体同人（服饰一致） 16×10＝160，皮肤同人（服饰一致） 11×20＝220，加权采样 A=500

• 10 epochs / 5000 步（＝A 500 × 10），batch 1，768 bucket

• LoRA rank 16 / alpha 16，LR 1e-4，AdamW + cosine（warmup 100，min 0.1）

• bf16 训练，fp16 导出（约 71 MiB）

• 单卡 RTX 4070 SUPER 12 GB，约 6 小时，最终 loss 0.2507

• caption 不含外貌词，身份由触发词吸收

Tequila 龙舌兰 (Arknights) · MOSS-Image Character LoRA

A character LoRA for Tequila from Arknights. Blonde hair, blue eyes, dog ears, dog tail; covers Base outfit (Elite 0), Base outfit (Elite 2), Cardwinner skin, Stage live skin. Trained on MOSS-Image Base; running it on the SFT model is recommended.

Known issue: Clothing details reproduce poorly, likely because each outfit receives too few total steps.

【Character trigger】

MoChaTequila

Put it on the first line of the positive prompt, exactly as written — no extra spaces or punctuation — then a newline and your scene.

The trigger alone already produces the character's features (blonde hair, blue eyes, dog ears, dog tail), but it does not specify the outfit — with no outfit tags the clothing is random, which is normal for a character LoRA.

【Base outfit (Elite 0)】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, hood, hooded jacket, white jacket, blue shirt, white shorts, yellow belt, sandals, sunglasses, necklace, sword, scabbard, sheath

【Base outfit (Elite 2)】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, hood, hooded jacket, white jacket, blue tank top, white shorts, sandals, sunglasses, necklace, sword, scabbard, sheath

【Cardwinner skin】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, grey jacket, white shirt, black pants, black gloves, suspenders, playing cards, sunglasses, black belt, black choker

【Stage live skin】

MoChaTequila, 1boy, blonde hair, blue eyes, dog ears, dog tail, black shirt, blue shirt, grey pants, two-tone gloves, sunglasses, microphone, microphone stand, stage

【Outfit swap】

Keep the trigger and describe the clothes in more detail; if you don't want the official outfit, put it in the negative:

Negative: hood, hooded jacket, white jacket, sunglasses, sword, scabbard

【Recommended settings】

• Base model: MOSS-Image SFT (int8_convrot or bf16); Base also works

• LoRA weight: 0.9–1.1, starting from 1.0

• Steps: 40–60, 50 recommended; CFG 3.0 (3.0–3.5)

• Sampler: euler + sgm_uniform

【Training parameters】

• Base model: MOSS-Image Base (tested on Base and SFT)

• Data: 31 images (repeats included) = official art 4×30 = 120 + base-outfit fan art 16×10 = 160 + skin fan art (outfit matches) 11×20 = 220; weighted sampling A = 500

• 10 epochs / 5000 steps (= A 500 × 10), batch 1, 768 bucket

• LoRA rank 16 / alpha 16, LR 1e-4, AdamW + cosine (warmup 100, min 0.1)

• bf16 training, fp16 export (~71 MiB)

• Single RTX 4070 SUPER 12 GB, ~6 h, final loss 0.2507

• captions contain no appearance words; identity is absorbed by the trigger