
触发词
文件清单 线路:镜像(hf-mirror.com) · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
SidelightGAL: Mya — Realistic Aisan Girl | MOSS-Image Character LoRA
作者:yinhades
A 22-year-old Aisan virtual character for MOSS-Image. One trigger word keeps her face consistent across hairstyles, outfits, poses and scenes. 一个 22 岁的亚洲虚拟角色,专为 MOSS-Image 训练。一个触发词,就能让她的脸在不同发型、服装、姿势和场景下保持一致。 What this is Mya is an identity LoRA for MOSS-Image. It was trained on 30 self-produced renders covering 8 angles (front, 45°, 90° on both sides, plus full-body front / 45° / 90°), a range of hairstyles and both everyday and body-emphasizing outfits. Trigger word: Mya-2Recommended settings (for Moody Moss Mix V4.1, the base it was trained on)
| Setting | Value |
|---|---|
| Sampler | Euler (flow-matching) |
| Steps | 25–30 |
| CFG | 3.0 |
| Shift | 5.0 |
| Skip layer | 22 |
| LoRA weight | 0.65–0.80 (0.70 is the sweet spot) |
| Resolution | 896×1152 / 1024×1024, multiples of 16 |
How to prompt it properly
- Always add the trigger: Mya-2, ...
- Describe the hairstyle immediately after the trigger. Hairstyle was written into every training caption, so hair follows your text while the face stays hers — you can freely switch between blunt bangs, middle part, high ponytail, low bun, top knot or a shoulder-length bob.
- Do not describe facial features (big eyes, small face, fair skin, etc.). The trigger already carries her face; adding descriptors drags it toward a generic "AI pretty face".
- Her figure is carried by the trigger too — you do not need to describe bust/waist/hip every time. Add shape words only when you want a specific emphasis. Example:
Mya-2, blunt bangs, dark brown hair in a high ponytail, close-up bust portrait, looking at the camera, light grey seamless background, soft key light with a rim light, realistic portrait photography
Mya-2, 中分,深棕色长直发披肩,无刘海。紧身黑色针织长裙,侧身四分之三站立,白色墙面,单侧窗光勾勒轮廓,七分身,写实摄影。
Training details
- 30 images at 1296×1728, self-produced with a consistent character
- LoRA rank 32 / alpha 32, LR 1.5e-4, 600 steps
- Hairstyle described in captions; facial features and body measurements deliberately left out so the trigger absorbs them Known limits
- Best on Moody Moss Mix V4.1 (the exact base).
- Text rendering (MOSS's signature strength) is unaffected by this LoRA. License / usage This is a fictional, fully synthetic character (AI-generated). No real person is depicted. Commercial use follows the base model's and platform's terms; please do not present it as a real person.
简介(中文)
这是什么 Mya 是一个面向 MOSS-Image 的人物身份 LoRA(人物一致性模型)。训练素材为 30 张自制渲染图,覆盖 8 个视角(正脸、左右 45°、左右 90°,以及全身正面 / 45° / 90°),并包含多种发型、日常与显身材两类服装。 触发词:Mya-2推荐参数(针对训练所用底模 Moody Moss Mix V4.1)
| 参数 | 数值 |
|---|---|
| 采样器 | Euler |
| 步数 | 25–30 |
| CFG | 3.0 |
| Shift | 5.0 |
| 跳层 | 22 |
| LoRA 权重 | 0.65–0.80(0.70 最稳) |
| 分辨率 | 896×1152 / 1024×1024,16 的倍数 |
提示词写法
- 永远加入触发词Mya-2, ...
- 发型写在触发词紧跟的第一句——发型写进了每一条训练标注,所以发型听你的文字、脸跟着触发词,齐刘海 / 中分 / 高马尾 / 低发髻 / 丸子头 / 齐肩短发都能自由切换
- 不要描写五官(大眼睛、小脸、皮肤白皙这类)。触发词已经承担了脸,再加描述会把脸拉向通用「AI 美人脸」
- 身材同样由触发词承担,不必每次重复描述胸腰臀;只有想强调某个部位时才加身材词 训练信息
- 44 张 1296×1728 素材,同一角色的自制渲染
- LoRA 秩 32 / 缩放 32,学习率 1.5e-4,600 步
- 标注里写发型、不写五官与身材,让触发词吸收这些特征 已知局限
- 在 Moody Moss Mix V4.1 上效果最好;
- 不影响 MOSS 招牌的文字渲染能力 使用声明 这是一个完全合成的虚构角色(AI 生成),不对应任何真实人物。商用请遵守底模与平台条款,请勿把它当作真人使用。
Tags
MOSS-Image, character, woman, chinese, realistic, photorealistic, identity, lora, portrait, Mya-2