1.1anima 做了数据集的重清洗，现在还原效果更好
V1.1: The Anima dataset has been thoroughly re-cleaned, resulting in improved style fidelity.
About Krea2 vision
Krea2 responds much better to clear scene direction than to a wall of traditional quality tags. Describe what the character is doing, where the camera is placed, how the scene is lit, what occupies the foreground and background, and what emotional impression the image should create.
For a classic official character illustration, try something like:
sen_kiseki_style, a full-body character illustration of a young female swordswoman standing on a clean white background, wearing an elegant red and black military uniform with gold details, holding a slender sword at her side, calm confident expression
For a cinematic scene, try:
sen_kiseki_style, a silver-haired swordsman standing alone on a rain-covered railway platform at night, viewed from a low camera angle, warm station lights reflected across the wet ground, his coat moving in the wind as a train disappears into the distance
You do not need to summon the ancient spirits of masterpiece, best quality, absurdres every time. Describe the image you actually want. Krea2 is surprisingly good at listening.
相比堆叠大量传统质量标签，Krea2 更适合清晰、具体的场景描述。建议直接说明：
人物正在做什么
镜头位于什么位置
场景使用怎样的光照
前景和背景分别有什么
希望画面传达怎样的情绪
如果想生成经典的官方人物立绘，可以尝试：
sen_kiseki_style, a full-body character illustration of a young female swordswoman standing on a clean white background, wearing an elegant red and black military uniform with gold details, holding a slender sword at her side, calm confident expression
如果想生成带有电影感的剧情场景，可以尝试：
sen_kiseki_style, a silver-haired swordsman standing alone on a rain-covered railway platform at night, viewed from a low camera angle, warm station lights reflected across the wet ground, his coat moving in the wind as a train disappears into the distance
About This LoRA
This is the Krea2 edition of my Murakami Fumika (村上文郁) style LoRA!
Murakami Fumika is one of the key artists behind the modern visual identity of Nihon Falcom. Her work is closely associated with titles such as the Trails of Cold Steel series, Trails into Reverie, Tokyo Xanadu, and, most recently, Kyoto Xanadu: The Blooming Phantom.
This LoRA aims to capture that unmistakable modern Falcom flavor: clean and elegant linework, beautifully shaped faces, distinctive eye rendering, elaborate costumes, polished character designs, and the feeling of an official JRPG illustration.
Note: This LoRA is for educational and personal sharing purposes only. Please do not use it for commercial purposes.
关于这个 LoRA
这是我的村上文郁（Murakami Fumika）画风 LoRA 的 Krea2 版本！
村上文郁是构成 Nihon Falcom 近年视觉风格的重要画师之一。她的工作与《闪之轨迹》系列、《创之轨迹》《东京迷城》等作品密切相关，在最新的《京都迷城：樱花幻舞》中也能看到这一视觉风格的延续。
本 LoRA 主要尝试捕捉现代 Falcom 美术中非常有辨识度的元素：干净优雅的线条、美型的人物面部、独特的眼睛塑造、精致复杂的服装设计，以及仿佛能够直接出现在日式 RPG 官方宣传图中的完成度
注：本模型仅供个人学习与交流分享，请勿用于商业用途。
Creator's Note: This is the very first LoRA I've ever created! There might be some minor bugs or quirks, but I really hope you enjoy using it. Also, this is an Illustrious-based model, so I highly recommend using Wai's base models for the best generation results.
这是我本人制作的第一个 LoRA！可能还会存在一些小 bug 或者不太稳定的地方，希望大家能够喜欢并多多包涵。另外，需要特别说明的是，这是一个**光辉系（Illustrious）**架构的模型，强烈推荐搭配 Wai 系列底模使用，出图效果最佳🤓
Description
About this LoRA:
This is the Anima-based version of my Murakami Fumika (村上文郁) style LoRA — same artist, fresh foundation. She is the main illustrator and character designer for Nihon Falcom's recent titles, including the Trails of Cold Steel (*Sen no Kiseki*) series, Trails into Reverie, and Tokyo Xanadu.
The trigger word and overall behavior are kept consistent with my previous Illustrious version, so if you've used the old LoRA, you can drop this one in with almost no prompt rewriting.
Why an Anima version? What's improved?
I retrained this LoRA on the brand-new Anima Base v1.0 (Cosmos-Predict2 architecture with a Qwen text encoder) and ran side-by-side tests against the Illustrious version using identical prompts. The Anima edition shows clear advantages in:
- Multi-character compositions — handles 2–3 characters in the same scene much more cleanly. Feature bleeding between characters is dramatically reduced, and each character keeps their own coherence in group shots.
- Dynamic poses and action — fight scenes, dramatic angles, jumping/running/swinging poses come out noticeably more natural and less stiff. The Illustrious version tended to lock into static "character design" poses; this one breathes.
- Prompt adherence — Anima's Qwen text encoder parses complex prompts far better. Long descriptions, multiple objects, scene composition instructions all land more reliably.
- Natural language support — you can now mix Danbooru tags with English sentences in the same prompt. Especially useful for atmosphere, lighting, and character interaction.
Recommended Parameters:
* Trigger Word: sen_kiseki_style
* LoRA Weight: 0.7 – 1.0 (Anima generally wants slightly higher weights than Illustrious; sweet spot around 0.8)
* **Base Model:** Anima Base v1.0 ⚠️ Do NOT use this LoRA on Illustrious / NoobAI / SDXL models — different architecture, completely incompatible.
* CFG Scale: 4 – 5 (Anima needs lower CFG than Illustrious — 7 will overcook the image)
* Sampler: er_sde (officially recommended for Anima) or euler_a
* Steps: 30 – 50
Prompting Tips:
For that classic Falcom official art look, the same tags still work:
simple background, white background, solo, character design, official art
For Anima specifically, follow the recommended prompt order:
[quality tags], 1girl/1boy, [character], [series], sen_kiseki_style, [general tags]. [optional natural language description for scene/atmosphere]
Feel free to mix tags and natural language — this is one of Anima's strengths over the Illustrious era.
Note: This LoRA is for educational and personal sharing purposes only. Please do not use it for commercial purposes.
Creator's Note: This is my first LoRA trained on the Anima base. Since Anima is a very new model and the training workflow is still being refined, there may be some quirks or bugs. Hope you enjoy it — generation shares, bug reports, and prompt examples all very welcome in the comments! :))
关于这个 LoRA:
这是我之前发布的村上文郁（Murakami Fumika）画风 LoRA 的 Anima 基模新版本——同一位画师,全新底子。她负责了《闪之轨迹》系列、《创之轨迹》以及《东京迷城》等 Nihon Falcom 近期作品的核心人物设计与立绘。
触发词与整体行为和光辉版保持一致,用过老版本的朋友可以无痛迁移,几乎不用改提示词写法。
为什么要做 Anima 版本?有什么优势?
我在全新发布的 Anima Base v1.0(基于 Cosmos-Predict2 架构 + Qwen 文本编码器)上重新训练了这个 LoRA,并用相同的 prompt 和光辉版本做了对比测试。Anima 版本在以下几个方面有明显提升:
- 多人构图 —— 处理同画面 2–3 个角色的能力强了一大截。角色之间的特征泄漏显著减少,即使在群像场景下每个角色也能保持独立性。
- 动作和动态姿势 —— 战斗场面、戏剧性视角、跳跃/奔跑/挥剑等动作类构图,出图显著更自然,不再像光辉版那样容易锁死在"立绘"式的标准姿势。
- 提示词跟随能力 —— Anima 的 Qwen 文本编码器对复杂 prompt 的解析能力是真的强。长描述、多物件、场景组合指令的响应都更可靠。
- 自然语言支持 —— 现在可以把 Danbooru tag 和英文长句混着写,描述氛围、光照、角色互动尤其好用。
推荐使用参数:
* 触发词: sen_kiseki_style
* 推荐权重 (Weight): 0.7 – 1.0(Anima 整体需要的权重比光辉略高,推荐 0.8 左右)
* **推荐底模:** Anima Base v1.0 ⚠️ 注意:本 LoRA 不能在光辉 / NoobAI / SDXL 系模型上使用,架构完全不同。
* CFG Scale: 4 – 5(Anima 需要的 CFG 比光辉低,7 会过曝)
* 采样器: er_sde(Anima 官方推荐)或 euler_a
* 步数: 30 – 50
出图小贴士:
想还原 Falcom 官方立绘风格,这些 tag 依然有效:
simple background, white background, solo, character design, official art
Anima 专属建议:按推荐 prompt 顺序排列:quality tags], 1girl/1boy, [character], [series], sen_kiseki_style, [general tags]. [optional natural language description for scene/atmosphere]
可以自由混合 tag 和自然语言——这是 Anima 相对光辉时代最大的进步之一。
注: 本模型仅供个人学习与交流分享,请勿用于商业用途。
作者的话: 这是我第一个 Anima 基模的 LoRA。Anima 模型刚发布不久,LoRA 训练流程也还在摸索中,可能还会存在一些小 bug 或不稳定的地方。欢迎大家下载试用,有任何 bug、prompt 示例或建议都可以在评论区告诉我:))
FAQ
Comments (2)
去年给界做mod的时候要是有这个lora，我就不至于自己拿笔画到腱鞘炎了效果还不太满意了23333
我好像还用过大佬你的mod，谢谢💕