📌 请务必花两分钟看完简介与说明，以获得最佳的使用体验！

(Please take two minutes to read the description and guide for the best experience!)

[CN] 是的，这是一个风格 LoRA，但在某种意义上，它也算是一个“全家桶”？

一个更比十个强？哈哈哈，开个玩笑啦！

不过还是请您至少耐心看完下方的使用说明哦~~~

[EN] Yes, this is an art style LoRA, but in a way, it's also a massive "All-in-One" character package.

Buy one, get ten free? Hahaha, just kidding!

But please do at least skim through the usage instructions below~~~

💬 作者碎碎念 (A Note from the Creator)

[CN] 顺便提一句，本人的提示词功底相对薄弱，上方展示的所有例图，仅仅是为了验证本 LoRA 的基础泛化功能与画面质感。

为了展现模型最原始、最纯粹的状态，所有的测试图除了本 LoRA 之外，均未加载任何其他的人物或画风 LoRA，也未在提示词中引用任何特定的画师风格。

更多的进阶玩法与绝美构图，就全权交由各位“艺术家”们自行开发与体验了。我非常期待能在评论区看到大家的惊艳返图！

[EN] As a quick side note, my prompting skills are relatively basic. All the showcase images provided above are purely to verify the baseline functionality, generalization, and visual texture of this LoRA.

To demonstrate the model in its most raw and unadulterated state, no other character or style LoRAs were loaded for these test images, nor were any specific artist names referenced in the prompts.

I leave the exploration of advanced techniques and breathtaking compositions entirely in the capable hands of you brilliant artists. I absolutely cannot wait to see the stunning artwork you share in the reviews!

📖 简介 (About This LoRA)

[CN] 极致的 Fortiche 工业美学，献给所有沉醉于底城霓虹与皮城荣光的创作者。

本 LoRA 旨在尽可能复现《双城之战》（第一部与第二部）中极具张力的 2.5D 厚涂与令人窒息的电影级光影美学。

无论你是想为心爱的角色重塑硬核的蒸汽朋克质感，还是想生成史诗般的宏大场景，这个模型都能为你提供降维打击般的视觉体验。

为了保证画风的绝对纯净度与质感，本模型的训练集 100% 来源于原剧的高清精选剧照，未掺杂任何同人或杂图。强烈推荐使用横图（Landscape）进行生成，以获得最佳的电影构图体验！

[EN] The ultimate Fortiche industrial aesthetics, dedicated to all creators captivated by the neon lights of Zaun and the glory of Piltover.

This LoRA is a humble attempt to replicate the highly dynamic 2.5D thick-painting art style and breathtaking cinematic lighting of Arcane (Seasons 1 & 2).

Whether you want to clothe your favorite characters in a hardcore steampunk texture or generate epic, monumental scenes, this model delivers a visually stunning experience.

To ensure absolute stylistic purity, the training dataset consists 100% of carefully curated high-resolution screenshots from the original series, with zero fanart or mixed styles. It is highly recommended to use landscape orientation for generation to achieve the best cinematic composition!

⚙️ 核心使用说明 (Usage Guide)

全局画风触发词 (Global Trigger): arcane artstyle

[CN] 强烈建议置于提示词最开头。

[EN] Highly recommended to place at the very beginning of the prompt.

人物/场景触发词 (Secondary Triggers):

[CN] 请见下方各角色与场景的专属英文触发词。

[EN] Please refer to the exclusive English trigger words for each character/scene below.

推荐画幅 (Aspect Ratio):

[CN] 强烈推荐横图（如 16:9, 3:2 等），以完美契合大片级别的电影构图。

[EN] Horizontal aspect ratios (e.g., 16:9, 3:2) are strongly recommended to perfectly match the cinematic framing.

✨ 稳定触发人物咒语 (Character Prompts)

💡 提示与用法 (Tips & Usage)

[CN] 提示：角色名旁括号内的英文即为该角色的“二级触发词 (Secondary Triggers)”。

下方列出的角色提示词已经过精简，仅保留了最核心的外貌特征。您直接复制粘贴即可精准“召唤”该角色，并且可以根据自己的创作需求，在提示词末尾自由添加动作描述或更换背景环境。

当然，如果您需要一个更完善的起手式，也非常欢迎直接复制我上方展示的例图中的完整提示词，来进行修改和“抽卡”。

[EN] Note: The English text in parentheses next to the character's name serves as their specific "Secondary Trigger".

The prompt snippets listed below have been streamlined to include only their core visual features. You can copy and paste them directly to accurately "summon" the character. Feel free to append any action descriptions or change the background settings at the end of the prompt according to your creative needs.

Alternatively, if you'd like a solid starting point, you are more than welcome to copy the full prompts from my uploaded showcase images and modify them as you see fit!

爆爆 (Powder): Powder, young child with short blue hair and freckles, wearing a green shirt, indoor blurry background.
爆爆 - IF线 (DreamPowder): DreamPowder, young girl with blue hair and freckles, wearing a choker and necklace, standing outdoors near a tree, portrait.
金克丝 (Jinx): Jinx, young woman with long braided blue hair, freckles, shoulder tattoo, brown choker, blurry background.
蔚奥莱 - 幼年 (Violet): Violet, young woman with short pink and red hair, wearing an open red jacket over a brown vest, outdoors during the day.
蔚 (ArcaneVi): ArcaneVi, woman with short asymmetrical red hair, facial tattoos, piercings, wearing a red jacket, blurry brown background.
凯特琳 (Caitlyn): Caitlyn, woman with long black hair, wearing a blue jacket, beret, and white ascot, black background.
梅尔 (MelMedarda): MelMedarda, dark-skinned woman with black dreadlocks bun, glowing facial tattoo, gold armor and jewelry, green background.
安蓓撒 (Ambessa): Ambessa, dark-skinned woman with short black dreadlocks, wearing gold armor with shoulder plates, portrait shot.
希尔科 (ArcaneSilco): ArcaneSilco, man with short slicked black hair, heterochromia red eye, facial scar, wearing a red shirt and jacket, black background.
塞薇卡 (Sevika): Sevika, dark-skinned woman with short black hair, mechanical prosthetic arm, wearing a jacket, graffiti alley at night.
杰斯 (Jayce): Jayce, mature man with short black hair and facial hair, wearing a white jacket over a black shirt, black background.
维克托 (Viktor): Viktor, dark-skinned man with short brown hair, wearing a brown collared shirt and vest, indoor setting.
黑默丁格 (Heimerdinger): Heimerdinger, yordle male with pointy ears and blonde hair, wearing a blue jacket, indoor brown background.
辛吉德 (ArcaneSinged): ArcaneSinged, bald man with facial scar and colored sclera, wearing a scarf, outdoors with blurry green background.
艾克 - 野火帮首领 (FirelightEkko): FirelightEkko, dark-skinned man with dreadlocks, wearing a brown jacket, cinematic lighting, portrait.
艾克 - IF线 (DreamEkko): DreamEkko, dark-skinned man with white dreadlocks tied up, wearing a green jacket over a white shirt, blurry background.

🎬 场景与特效指令 (Environments & VFX)

[CN] 提示：已提炼最核心的视觉标签，可直接搭配全局画风词使用。

[EN] Note: Core visual tags have been extracted. Use them directly alongside the global style trigger.

祖安底城 (Zaun Scenery): Zaun_scenery, cinematic night scene, weathered iron bridge, dense industrial buildings, flickering neon lights, wet cobblestone street, toxic emerald smog.

皮尔特沃夫 (Piltover Scenery): Piltover cityscape, outdoors during the day, large airship flying above fantasy-style buildings and towers, bright blue sky, no human.

双城爆炸特效 (Explosion VFX): arcane vfx, explosion, no human.

[CN] 温馨提示：本 LoRA 的 2600+ 张训练集全部提取自电影级原生宽银幕素材。虽然上方展示的例图为了适配网页排版全部采用了竖图，但本模型在生成【横图（Landscape/Cinematic aspect ratio）】时会表现得更加游刃有余、质感拉满。强烈推荐各位艺术家尝试横图构图！

[EN] Pro-tip: This LoRA was trained on over 2,600 frames of native, cinematic widescreen material. While the showcase gallery uses portrait images optimized for web scrolling, this model is inherently more comfortable and visually stunning when rendering in [Widescreen/Landscape aspect ratios]. I highly encourage you brilliant artists to unleash its full potential in cinematic widescreen formats!

🧪 炼丹纪实与理论探索 (Training Insights & Theoretical Attempts)

[CN] 万里挑一的地狱级收集：这次《双城之战》模型的炼制，对我而言是一场极度消耗精力的持久战。初始素材来源于两部剧集中每两秒一次的自动截图，总计 20,000 多张原始帧。我从中手动初筛出 5,000 张进行细致分类，最终以极其苛刻的标准，精选出 2,639 张顶级剧照作为核心训练集。

说句心里话，如果没有极大的决心和耐心，请千万不要轻易尝试这种大型训练集的精细化收集与打标。我必须承认，我严重低估了构建《双城之战》训练集的工作量。它的收集与分类难度远超我之前处理的《紫罗兰永恒花园》。这不仅是因为剧中出场且有名字的角色繁多，更因为皮尔特沃夫与祖安在建筑风格和画面色调上的巨大差异，需要进行极其繁琐的细分。

最让人崩溃的是，第一部的整体画面风格偏暗。在电脑的文件夹缩略图里，根本看不清图片的内容与细节，很大一部分图片必须单张点开放大才能辨认，分类工作堪称地狱级考验。

当我意识到这个“天坑”的时候，我已经分好几百张图了——总不能半途而废吧！最后完全是靠着“来都来了，干都干了”的执念，硬生生熬完了整个训练集的构建。

单单是前期的收集与分类，实际工作时长就远超 40 个小时。直到昨天，模型才被炼制出来，并且进行了测试，挑选出发挥比较稳定的其中一个。

此外，为了防止最终出图“偏科”、导致模型生成的画面永远是一团死黑，我又咬紧牙关，补充收集了画面光影稍亮一些的第二部剧照，来平衡整体的数据分布。

[EN] A Hellish, One-in-a-Million Curation Process: Creating this Arcane model was an absolute marathon. The initial source material came from automated screenshots taken every two seconds across both seasons, yielding over 20,000 raw frames. I manually filtered this down to 5,000 for detailed categorization, and finally, with exceedingly strict standards, handpicked a core dataset of 2,639 top-tier cinematic shots.

To be completely honest, unless you have immense determination and patience, I strongly advise against attempting the meticulous collection and tagging of a large-scale dataset like this. I must admit, I severely underestimated the workload. Building the Arcane dataset was exponentially harder than my previous work on Violet Evergarden. Not only is the cast of named characters much larger, but the stark architectural and visual contrasts between Piltover and Zaun required incredibly tedious sub-categorization.

The most soul-crushing part? Season 1's notoriously dark art style. In standard folder thumbnails, it was practically impossible to make out the contents or details. A huge portion of the images had to be opened and zoomed in individually just to identify them, turning the sorting process into a hellish ordeal.

By the time I realized the sheer scale of this nightmare, I had already sorted hundreds of images. I couldn't just give up halfway! Ultimately, it was pure sunk-cost stubbornness—the "I'm already in too deep, might as well finish it" mentality—that forced me to power through.

The initial gathering and sorting phase alone consumed well over 40 hours of actual labor. It was not until yesterday that the model was trained and tested, and one that performed consistently was selected.

Furthermore, to prevent the model from overfitting to dark tones and constantly producing muddy, pitch-black generations, I bit the bullet and gathered brighter shots from Season 2 to balance out the overall lighting data.

🎯 针对“二级触发词”的底层验证与创新尝试 (Validating "Secondary Triggers" & Innovative Attempts)

[CN] 一次“画风与人物融合”的创新探索：我每次炼制 LoRA 时，都会尽量给自己找一点“创新点”。

这次《双城》模型的核心探索，就是验证在庞大的全局画风触发词统摄下，能否通过“多重二级触发词”来稳定召唤特定人物的不同形态。从底层逻辑来看，在 AI 的视野里并没有绝对的“画风 LoRA”与“人物 LoRA”之分。

当我们用海量图片绑定一个全局词时，AI 学习到的是共性的笔触和光影；而当我们在此基础上，为特定人物打上专属标签时，AI 理论上就能在全局画风的框架内提取并记住这个人物。这就是在画风库里“顺便”练出人物的基础。

[EN] An Innovative Exploration of Merging Style and Characters: Every time I train a LoRA, I try to find a little "innovation point" for myself.

The core exploration of this Arcane model was to verify whether specific characters in their various forms could be stably summoned using "multiple secondary trigger words" under a massive global style trigger. From an underlying logic perspective, AI doesn't strictly distinguish between a "Style LoRA" and a "Character LoRA."

When we bind massive amounts of images to a global tag, the AI learns common brushstrokes and lighting; if we assign specific tags to characters within that pool, the AI theoretically memorizes them within the stylistic framework. This is the foundation of "conveniently" training characters within a style pool.

[CN] 实验结果（成功的验证）：这次的尝试可以说是“喜忧参半”。成功的一面在于，二级触发词的隔离效果非常惊艳！

我将艾克的素材进行了细分，结果仅需极简的描述，FirelightEkko（野火帮首领）和 DreamEkko（IF线穿越时空）就能被完美区分开来。

同样的，Powder（爆爆）、DreamPowder（IF线爆爆）和 Jinx（金克丝），以及幼年 Violet（蔚奥莱）与成年 ArcaneVi（蔚），都被各自的二级触发词精准隔离。

当然，这也是因为像金克丝（149张）、蔚（132张）、凯特琳（119张）不仅素材量充足，而且她们本身就是大模型底模自带认知的基础角色，所以召唤起来十分轻松。

[EN] The Results (Successful Validations): The results were a mix of success and failure. On the successful side, the isolation effect of secondary trigger words was astonishing!

I subdivided Ekko's dataset, and as a result, FirelightEkko (Firelight Leader) and DreamEkko (IF-line Dream) can be perfectly distinguished with minimal prompts.

Similarly, Powder, DreamPowder, and Jinx, as well as young Violet and adult ArcaneVi, are flawlessly isolated by their respective secondary trigger words.

Naturally, this is also because core characters like Jinx (149 images), Vi (132 images), and Caitlyn (119 images) not only had ample dataset images but also already had base recognition within the foundational models, making them incredibly easy to summon.

[CN] 实验结果（失败的反思与权重淹没）：失败的一面则进一步印证了我对“权重稀释”的推测。更确切地说，这些非剧情核心角色并非“完全无法触发”，而是根据训练集素材数量与质量的差异，产生了不同程度的“召唤困难症”。

在接近 3000 张的庞大画风底盘下，普通的单人素材及格线显得微不足道，最直观的表现就是“神似而形不准”：

范德尔 (Vander)：虽然能被唤醒，但脸部无法稳定达到与原著高度相像的程度。

史密奇 (Smeech)：脸部细节严重丢失。不过确实还是丑得一比，我随便测了几张就直接放弃了（笑~）。

沃里克 (Warwick)：虽然已经是纯粹的非人怪物形态，但由于受到海量人类素材的权重污染，生成时依然会隐隐约约带上人类的五官轮廓。

至于凯特琳的母亲 (Cassandra)、洛里斯 (Loris) 和麦迪 (Maddie - 云顶之弈棋子)、怀斯 (Wyatt - 艾克的父亲) 以及“小艾克”等人，由于训练集数量实在太少，效果更是惨不忍睹。

这也引发了我的一个思考：如果后续加大针对这些次要人物的搜集力度，适当放宽同类图片的筛选质量，尝试“用数量来弥补质量的不足”，是不是就有大概率能把他们从底盘权重中“拽”出来？

我想，这也将是我日后填坑、优化这个 LoRA 的一个明确方向。因此，现阶段这些“沉睡”角色的二级触发词我就先不放出来了，放了也没有意义。

[EN] The Results (Failed Reflections & Weight Dilution): The failure side further confirmed my theory on "weight dilution." To be precise, these non-core characters are not "completely untriggerable," but rather suffer from varying degrees of "summoning difficulty" depending on the quantity and quality of their training images.

Under the massive 3,000-image style base, the passing grade for standard single-character datasets becomes insignificant. The most direct manifestation is that they "lack accurate likeness":

Vander: He can be triggered, but his face cannot stably resemble his original design to a high degree.

Smeech: Severe loss of facial details. However, he still indeed looks ugly as hell—I tested it a few times and just gave up, lol.

Warwick: Even though he is a pure non-human monster, the generations still vaguely incorporate human facial features, likely contaminated by the massive human dataset.

As for Caitlyn's mother (Cassandra), Loris, Maddie (from TFT), Wyatt (Ekko's father), and "Little Ekko," their results are even worse due to the extremely limited training images.

This also sparked a hypothesis: If I increase the data collection efforts for these secondary characters in the future and slightly relax the quality standards for their images (using quantity to compensate for quality), is it highly likely they could be "pulled out" from the base weight?

I think this will be a clear direction for optimizing this LoRA in the future. Therefore, for now, I won't list the secondary trigger words for these "slumbering" characters, as it would be meaningless.

[CN] 一点碎碎念：说句心里话，我所有的炼丹方式和底层逻辑，完全是在和 AI（Gemini）的一次次交流探讨中摸索学来的。而这种“多重二级触发词”的体系，也是我脑洞大开想出来的一个小假设。

这次能够切实验证它的可行性，算是我做这个《双城》模型得到的最大收获了。

当然啦，也可能是我鼠目寸光了，也许这个小实验在各位炼丹大佬们的眼里早就不是什么秘密了，哈哈。就当是我分享的一点个人心得吧！

[EN] A Little Personal Note: To be completely honest, all my training methods and understanding of underlying logic were learned entirely through continuous conversations and explorations with AI (Gemini). The concept of this "multiple secondary trigger word" system was just a little hypothesis I brainstormed.

Being able to practically verify its feasibility is my biggest harvest from making this Arcane model.

Of course, I might just be a frog in the well, and this little experiment might be common knowledge to veteran model trainers, haha. Just consider this my personal training diary and insights!

🛡️ 给同好的避坑指南与推测 (Tips & Observations for Fellow Creators)

[CN] 人物权重的“唤醒及格线”：结合我之前制作《紫罗兰永恒花园》的经验（1300 多张底图中有近 500 张薇尔莉特单人图，出图细节极其惊艳），我这次严重低估了画风大模型下的人物权重稀释效应。通常炼制单个人物只需 20 到 120 张素材；但在接近 3000 张的画风大盘里，这点素材简直是杯水车薪。

我个人推测：在庞大的画风库中，想要靠二级触发词稳定召唤某个角色，单人素材绝对不建议低于 80 张，最好能有 150 到 300 张涵盖各种角度的高质量图片。比例似乎是个重要参考——也许总集大小的 5% 到 10% 就是相对稳定的唤醒阈值。

这也解释了我的遗憾：我最初创建了四十几个角色文件夹，但对于范德尔、伊莎、小艾克、麦罗等素材量不达标的角色，他们在庞大的画风权重下陷入了“沉睡”，目前无法稳定触发。放出的仅是能经受住考验的核心人物。

[EN] The Threshold for Character Awakening: Drawing from my Violet Evergarden LoRA (where Violet had nearly 500 images out of 1,300, resulting in breathtaking details), I severely underestimated the dilution of character weights in a massive style model. While 20 to 120 images are enough for a standalone character, it's a drop in the ocean within a 3,000-image dataset.

I estimate that to stably summon a character using a secondary trigger word in a large-scale style pool, you should not have fewer than 80 high-quality images, and ideally need 150 to 300 images covering various angles. The ratio seems crucial—perhaps 5% to 10% of the total dataset is the threshold for awakening a character.

This explains my regret: I initially created over 40 character folders. Characters who didn't meet this threshold (like Vander, Isha, Little Ekko, Mylo) were kept as style base but fell "asleep" under the massive weight and cannot be stably triggered. Only the core characters who survived the weight test are provided.

[CN] 关于素材筛选的心得：如果在数千张规模的训练集里保持统一且严格的选拔标准，这些积少成多的细节会在生成质量上产生质变。

在删图时尽量严格，连续帧和模糊帧请精准清除。

举个例子：如果同一场景人物因为微小动作同时产生了正脸和侧颜，我建议优先保留侧颜。正脸相对容易收集，而侧面透视包含的体积信息往往更大，能更好地帮助模型学习空间感。

[EN] Thoughts on Image Curation: Maintaining a consistent and rigorous selection standard across thousands of images will lead to a qualitative leap in your generations. Be incredibly strict when deleting: precisely remove continuous and blurry frames.

For example, if a scene provides both a front-view and a side-profile due to a slight movement, I recommend prioritizing the side profile. Front views are easy to collect, whereas side profiles contain more volumetric information, better helping the model learn spatial depth.

🛠️ 推荐工具链与软件说明 (Recommended Tools & Software Guide)

[CN] 为了方便各位同好复刻或制作自己的大型训练集，以下是我在整个流程（从两万张截图到最终精选裁剪）中所使用的全套免费软件。如果你对这些工具的具体操作有疑问，可以直接向 Gemini 提问获取详细教程：

自动截图： PotPlayer（用于从剧集中每两秒自动截取高画质底图）

筛选图片： FastStone Image Viewer（非常适合用于快速浏览、标记并删除连续帧/模糊帧）

裁剪缩小图片： XnConvert（强大的批处理工具，用于统一裁剪构图并调整图片分辨率）

高清修复： Video2X（优秀的开源 AI 放大工具，用于提升部分低分辨率素材的画质）

缝合视频： LosslessCut（极速无损的视频剪切与合并工具，用于前期快速筛选高质剧集片段）

[EN] To help fellow creators replicate or build their own large-scale datasets, here is the complete pipeline of free software I used throughout the entire process (from 20k screenshots to final selection and cropping). If you have any questions about how to use these tools, you can ask Gemini directly for detailed tutorials:

Auto Screenshot: PotPlayer (For automatically extracting high-quality base frames from episodes every 2 seconds).

Image Curation: FastStone Image Viewer (Excellent for fast browsing, tagging, and ruthlessly deleting continuous or blurry frames).

Image Cropping & Resizing: XnConvert (A powerful batch-processing tool to unify compositions and adjust resolution).

AI Upscaling: Video2X (A superb open-source AI upscaler used to enhance the quality of certain low-res source materials).

Video Stitching: LosslessCut (An ultra-fast, lossless video cutting and merging tool used to quickly isolate high-quality episode segments initially).

💌 遗憾、未来计划与一点小小的期盼 (Regrets, Future Plans & A Small Wish)

[CN] 关于遗憾与休息：坦白地说，无论是之前的《千与千寻》、《紫罗兰永恒花园》，还是这次的《双城之战》，由于各项要素的比例实在难以完美平衡，它们都没能完全达到我心目中十全十美的标准，依旧存在着明确的优化空间。

由于个人的时间与精力终究有限，这次《双城》庞大的收集与分类工作确实让我感到十分疲惫。

因此，在接下来的一段时间里，我可能会稍微放慢脚步，先发布一些以前积攒的单人角色 LoRA 当作休息。

[EN] On Regrets & Taking a Break: To be completely honest, whether it was Spirited Away, Violet Evergarden, or now Arcane, none have reached my absolute standard of perfection due to the immense difficulty of balancing all element ratios. There is still clear room for optimization.

Since a single person's time and energy are ultimately limited, the massive workload of Arcane has genuinely left me exhausted. Therefore, I plan to slow down for a while and release some previously made single-character LoRAs just to take a breather.

[CN] 关于填坑与期盼：如果大家不嫌弃，觉得这个模型用着还算顺手，后续等我恢复了精力，我会尝试进一步优化训练集。看看能不能把那些素材量较少的角色也给“抢救”出来（希望能弥补那 48 个角色文件夹的遗憾），并进一步提升所有人物的换装灵活性与微表情张力。

最后，如果这个模型偶尔能为您带来一点点绝妙的灵感，希望能得到您的一个点赞（Like）。当然，如果您愿意在评论区分享您跑出的绝美返图，那就足够让我无比满足和开心了！

你们的作品和反馈，永远是我后续“填坑”更新的最大动力。再次感谢大家一直以来的包容与支持！

[EN] On Future Updates & Support: If you don't mind its flaws and find this model somewhat useful, I will try to further optimize the dataset once I have recovered my energy. I hope to "rescue" those characters with fewer source images (to fulfill the regrets of those 48 folders) and improve the flexibility of outfits and facial expressions for everyone.

Finally, if this model brings you even a spark of inspiration, a simple Like would mean the world to me. And if you are willing to share your beautiful creations in the reviews, that alone would make me incredibly happy and satisfied!

Your artwork and feedback are always my greatest motivation to keep updating. Thank you all again for your continued patience and support!

✨ 写在最后的真心话 (One Last Thing...)

[CN] 哈哈哈——开个玩笑啦！《双城之战》的坑，我以后肯定会填的。

至少，我会把“IF线（假设线）”的人物全部再次优化补齐：完好无损的希尔科、爆爆和蔚的母亲、艾克的父亲怀斯，还有那个没那么忧郁的范德尔……

唉，坦白讲，就算最后没人玩这个模型，我也一定会把它补全。

毕竟，难得艾克在剧里终于找到了那样一条完美的时间线，我要是不给这个 LoRA 配上一个完美的“大结局”，实在有些说不过去，对吧？

不过，既然后续要对 LoRA 进行大版本升级，我肯定需要对训练集进行更加严苛、细致的收集与配比。所以，下一次的更新大概要间隔相当长的一段时间了。

不知不觉竟然写了这么长的一段“小作文”，还请耐心看到这里的各位“艺术家们”多多海涵，再次感谢大家一直以来的包容与支持！

[EN] Hahaha—just kidding! I will absolutely continue updating this Arcane LoRA in the future.

At the very least, I plan to fully optimize and include all the characters from the "What-If" (IF) timeline: an unscarred Silco, Powder and Vi's mother, Ekko's father Wyatt, and a much less depressed Vander...

Honestly, even if no one ends up using this model, I will definitely complete it.

After all, considering Ekko finally found such a perfect timeline in the show, it just wouldn't feel right if I didn't give this LoRA a perfect "ending" too, right? However, since the next update will be a major upgrade, I will need to be even more meticulous with the data collection and ratio balancing. Therefore, it might take quite a long time before the next version drops.

Without realizing it, I've written a massive "essay" here. To all the "artists" who have read this far, thank you for bearing with me, and thank you all again for your continued patience and support!

❗ 免责声明 (Disclaimer)

[CN] 本模型仅供 AI 绘画爱好者学习、交流与同好分享使用，严禁用于任何形式的商业用途。本模型旨在向 Fortiche Production 卓越的美术风格致敬，无任何恶意侵权意图。《双城之战》的所有相关角色设计、美术风格及知识产权均归 Fortiche Production、Riot Games 及 Netflix 所有。使用者由此产生的一切商业纠纷或法律责任由使用者自行承担，模型作者概不负责。

[EN] This model is strictly for educational purposes, personal communication, and sharing among fans. Any form of commercial use is strictly prohibited. This model is created as a tribute to the exceptional art style of Fortiche Production, with absolutely no malicious intent to infringe upon copyrights. All character designs, art styles, and related intellectual properties of Arcane belong to Fortiche Production, Riot Games, and Netflix. The creator of this model assumes no responsibility for any commercial disputes or legal issues arising from its use.

Files with SHA256: 5a6e1872e1626619f8354d52a2b7be789c9454598a80f103aa60197630e9fd21