苏连雁 Sulianyan

《哀鸿：城破十日记》角色苏怜烟（苏连雁）的角色 LoRA。

本页面提供 Anima、Krea 2、MOSS 三个版本，请选择与你使用的底模对应的文件。

三个版本的角色专属触发词均为：

sulianyan

角色的主要特征为黑色长发、灰色眼睛、紫色花饰和垂落的珠串。各版本的提示词写法、推荐参数有所不同，下面分别说明。

Anima 版本

本版本基于 Anima系列底模训练。

推荐 LoRA 权重：0.80

建议测试范围：0.70–1.00

如果角色特征不明显，可以提高到0.90–1.00。

如果脸部、发型或原设服装绑定过强，可以降低到0.70–0.75。

基础提示词：

1girl, sulianyan, solo, black hair, very long hair, grey eyes, small pupils, parted bangs, hair ornament, purple flower, hair tassel

原服饰＋团扇：

1girl, sulianyan, solo, black hair, very long hair, grey eyes, small pupils, parted bangs, hair ornament, purple flower, hair tassel, hanfu, white robe, purple floral print, dark blue skirt, wide sleeves, holding fan, round fan, standing, looking at viewer

原服饰、不持团扇：

1girl, sulianyan, solo, black hair, very long hair, grey eyes, small pupils, parted bangs, hair ornament, purple flower, hair tassel, hanfu, white robe, purple floral print, dark blue skirt, wide sleeves, standing, looking at viewer, empty hands

如仍然出现团扇，可在负面提示中加入：

holding fan, round fan

非原服饰、不持团扇：

1girl, sulianyan, solo, black hair, very long hair, grey eyes, small pupils, parted bangs, hair ornament, purple flower, hair tassel, modern clothes, white blouse, black jacket, black skirt, standing, looking at viewer, empty hands

换装时如出现汉服残留，可在负面提示中加入：

hanfu, white robe, purple floral print, dark blue skirt, wide sleeves

防止眼睛颜色漂移，正面提示建议保留：

grey eyes, small pupils

Krea 2 版本

本版本基于 Krea 2 Raw 训练，使用 2000步检查点，rank／alpha 32／32。

推荐 LoRA 权重：0.80

换装、换发型时，如果原设影响过强，可以尝试降低到0.60。降低权重后，角色特征和紫花珠串也可能随之减弱。

Krea 2版本建议使用英文自然语言描述，把人物外貌、服装、动作和场景写清楚。

基础提示词：

sulianyan, an adult woman with gray eyes and small pupils, long black hair in a half-up hairstyle with parted bangs, and a violet flower hair ornament with dangling bead strands.

原服饰＋团扇：

sulianyan, an adult woman with gray eyes and small pupils, long black hair in a half-up hairstyle with parted bangs, and a violet flower hair ornament with dangling bead strands. She wears a moon-white wide-sleeved hanfu robe with deep navy edging and pale lavender embroidery, a dark purple pleated skirt, silver chest chains, and black cloth shoes. She stands beside a wooden lattice window, holding a round silk hand fan and looking toward the viewer with a calm expression. Soft warm window light, delicate 2D anime illustration, natural skin tones, carefully drawn fabric folds.

原服饰、不持团扇：

sulianyan, an adult woman with gray eyes and small pupils, long black hair in a half-up hairstyle with parted bangs, and a violet flower hair ornament with dangling bead strands. She wears a moon-white wide-sleeved hanfu robe with deep navy edging and pale lavender embroidery, a dark purple pleated skirt, silver chest chains, and black cloth shoes. She stands beside a wooden lattice window, looking toward the viewer with a calm expression. Her empty hands rest naturally in front of her waist. Soft warm window light, delicate 2D anime illustration, natural skin tones, carefully drawn fabric folds.

非原服饰、不持团扇：

sulianyan, an adult woman with gray eyes and small pupils, long black hair in a half-up hairstyle with parted bangs, and a violet flower hair ornament with dangling bead strands. She wears a white blouse beneath a fitted black jacket, a black skirt, and black shoes. She stands beside a bookstore window, looking toward the viewer with a slight smile. One empty hand rests on the window ledge, while the other hangs naturally at her side. Soft afternoon light, a gently blurred background, delicate 2D anime illustration.

换装提示：

直接替换整段服装描述，保留希望延续的人物特征。建议明确写出衣服的颜色、领型、袖型和主要结构，避免同时保留汉服与现代服装两套描述。

防止眼睛颜色漂移，建议保留：

gray eyes and small pupils

已测试的展示图设置：

• 底模：Pornmaster Krea2 v25 Turbo

• LoRA强度：0.80

• Sampler／Scheduler：ER_SDE／normal

• Steps／CFG：10／1

• 文本编码器：Qwen3-VL heretic int8

• VAE：Qwen Image VAE

这套10步、CFG1的设置对应上述Turbo底模，其他Krea 2底模请按各自的推荐参数使用。CFG1下不能依靠负面提示词消除服装、发饰或手部问题。

无发饰提示仍可能残留紫花或珠串，现代服装也可能带入花纹、盘扣等原设元素。精细眼部、复杂手指及镜像结构不保证每次一致。

MOSS 版本

本版本基于 MOSS-Image Base 训练，使用 第20轮／600步检查点，rank／alpha 32／32。

推荐 LoRA 权重：0.80

已测试0.60、0.80和1.00。建议先从0.80开始：

• 0.60会减弱角色影响，但部分近景会漏掉紫花。

• 1.00不一定更好，眼色、头发和衣装细节也可能变得更强。

• 不必为了增加角色特征直接把权重拉满，先补充明确的人物描述。

MOSS版本已在 Base和SFT 上进行实际出图。支持标签、自然语言及MOSS JSON Layout提示词。

基础提示词：

sulianyan, 1girl, solo, gray eyes, black hair, long hair, half updo, parted bangs, fair skin, purple flower hair ornament, dangling bead strands, a composed expression

原服饰＋团扇：

sulianyan, 1girl, solo, gray eyes, black hair, long hair, half updo, parted bangs, fair skin, purple flower hair ornament, dangling bead strands, hanfu, silver-gray robe, navy blue trim, long sleeves, wide sleeves, purple skirt, long skirt, pleated skirt, floral embroidery, holding a round silk hand fan, standing, looking at viewer, soft warm window light, delicate clean line art, softly blended shading

原服饰、不持团扇：

sulianyan, 1girl, solo, gray eyes, black hair, long hair, half updo, parted bangs, fair skin, purple flower hair ornament, dangling bead strands, hanfu, silver-gray robe, navy blue trim, long sleeves, wide sleeves, purple skirt, long skirt, pleated skirt, floral embroidery, standing beside a wooden lattice window, looking at viewer, one hand lightly gathering a sleeve at the waist, the other resting naturally on the window ledge, soft warm evening light, delicate clean line art, softly blended shading

非原服饰、不持团扇：

sulianyan, 1girl, solo, gray eyes, black hair, long hair, half updo, parted bangs, fair skin, purple flower hair ornament, dangling bead strands, white silk blouse, black high-waisted trousers, standing beside a tall window, looking at viewer, a very slight gentle smile, empty hands, soft warm side light, natural skin tones, a softly blurred low-contrast background, delicate clean line art, softly blended shading

换装时，移除以下原服装词，再加入新服饰：

hanfu, silver-gray robe, navy blue trim, wide sleeves, purple skirt, pleated skirt

如仍然出现团扇，可以尝试在负面提示中加入：

holding fan, round fan, hand fan

眼部与面部提示：

建议明确保留：

gray eyes

近景需要强调小瞳孔时，可以追加：

small pupils

如果脸部显得过于成熟，可以尝试补充：

soft youthful adult facial features, gently rounded cheeks, a natural gentle expression, minimal makeup

已测试的展示图设置：

• 底模：MOSS-Image SFT int8 convrot

• LoRA强度：0.80

• Sampler／Scheduler：Euler／simple

• Steps／CFG：50／3

• Shift：5

• SLG：single layer 22，0～1全程

• 分辨率：1024×1536

• 部分展示图使用SeedVR2放大至2560×3840

训练使用Base，以上是SFT展示图的实际设置。更换底模或文本编码器后，脸部、画风和颜色表现可能有所变化。

关于JSON Layout：

将触发词和人物外貌写进对应人物区域的caption，服装、动作、场景及光线也在JSON内部说明。使用工作流的“JSON直通”输入，避免在JSON外继续拼接标签或触发词。

已知限制：

紫花、珠串、眼色、衣扣及衣装纹样并不保证每张完全一致。去除发饰、严格纯黑白和纯复古印刷风仍有不稳定情况；复杂手部与小饰品需要逐图检查。

使用说明

本模型主要用于AI绘图测试、角色还原、学习交流及非商业同人创作。请在合理、合法的范围内使用，并尊重原作品及相关创作者的权益。

三个版本是分别训练的角色LoRA，请使用对应模型文件和提示词，不要将不同底模系列的权重混用。

模型主要学习角色相似度。受训练数据影响，某些服装、发饰、姿势、表情或背景元素可能更容易出现。如果角色显得过于固定，可以适当降低LoRA权重，并把希望改变的内容写清楚。

Krea 2版本还需遵守对应的 Krea 2 Community License 及其中引用的 Acceptable Use Policy。 各版本的使用范围同时受对应底模许可及页面标注权限约束。

免责声明

本LoRA为非官方同人模型，与原作者、画师、开发商、发行方及版权方均无任何官方关联。

模型涉及的角色形象、原作设定、美术素材及相关知识产权均归各自权利方所有。上传者仅进行了个人兴趣性质的模型训练与分享，不主张对原角色或原作品拥有任何权利。

请勿使用本模型冒充官方内容、进行虚假署名、误导性宣传，或从事可能损害原作者及相关权利方权益的行为。

如原作者、版权方或其他相关权利人对本模型的发布存在异议，可以联系我，我会认真处理相关问题。

欢迎返图，也欢迎提供角色还原、换装或画风搭配方面的反馈。反馈时附上模型版本、底模、完整提示词、LoRA强度和生成参数，会更方便定位问题。

This is a fan-made and unofficial character LoRA for Su Lianyan from 哀鸿：城破十日记, mainly created for AI image generation experiments, learning, and fan works.

Separate versions are provided for Anima, Krea 2, and MOSS-Image. Please use the version that matches your base model. The trigger word for all three versions is:

sulianyan

Start at 0.8 LoRA strength and refer to the settings listed for each version. The Krea 2 version was trained on Krea 2 Raw and uses the 2000-step checkpoint. The MOSS version was trained on MOSS-Image Base and uses the epoch-20 / 600-step checkpoint; examples have also been generated with MOSS-Image SFT.

I do not own the original character, artwork, source material, or any related intellectual property. All rights belong to their respective owners. This model is not affiliated with, endorsed by, or officially connected to the original creators, artists, publishers, developers, or copyright holders.

Please do not use this LoRA to impersonate official artwork, falsely claim authorship, mislead others, or infringe upon the rights of the original creators.

This model is mainly trained for character likeness. Because of the training dataset, certain outfits, poses, expressions, accessories, or background elements may appear more frequently than expected. If the character looks too fixed or the LoRA has too much influence on the image, try lowering its strength.

Use is also subject to the applicable base-model licenses. The Krea 2 version is a community derivative of Krea 2 Raw and is subject to the Krea 2 Community License Agreement and its incorporated Acceptable Use Policy. It is not an official Krea product or endorsed by Krea.

I am still learning LoRA training, so the model may not be perfect. Feedback, suggestions, testing results, and example images are always welcome.