<h1 id="arcane-style-lora-moss-image">Arcane Style LoRA — MOSS-Image</h1><h2 id="english">English</h2><h3 id="what-this-is">What this is</h3>Arcane Style is a style LoRA for MOSS-Image, trained to reproduce the painterly, high-contrast look of the Arcane animated series: thick oil-brush texture, hard-edged colour blocking, saturated violet / teal / amber palettes, and cinematic poster composition.It is a style LoRA, not a character LoRA. It does not lock a face — use it on any subject and it paints them in this language.<h3 id="important:-use-it-on-the-base-model">Important: use it on the Base model</h3>This LoRA was trained on MOSS-Image Base. It performs noticeably better on Base than on SFT — SFT has been aesthetic-tuned, and that post-training dilutes the LoRA's style signal. Train on Base, use it on Base.<h3 id="two-ways-to-run-it-and-i-recommend-the-fast-one">Two ways to run it — and I recommend the fast one</h3>There are two working configurations. Both give good results; the second is far faster and loses very little quality.① Default — 50 steps, no speed LoRA. The straightforward path. Nothing extra to download.② Recommended — 8 steps with the DMD speed LoRA. Pair this LoRA with the 8-step distillation of the Base model:<a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd">lodestones/MOSS-Image-Base-8-steps-dmd</a>It stacks cleanly on top of this LoRA. A 2.5 MP image renders in roughly 12 seconds, with only a very small quality difference — this is how every sample here was made.<h3 id="recommended-parameters">Recommended parameters</h3>Config ① — default (50 steps)<ul><li>Base model: MOSS-Image Base (not SFT)</li><li>Sampler: euler</li><li>Scheduler: simple</li><li>Steps: 50</li><li>CFG: 3.0 – 4.0</li><li>Shift: 5.0</li></ul>Config ② — recommended (8 steps, with the speed LoRA)<ul><li>Base model: MOSS-Image Base (not SFT)</li><li>Sampler: euler</li><li>Scheduler: simple</li><li>Steps: 8</li><li>CFG: 1.0</li><li>Shift: 5.0</li><li>Speed LoRA: <a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd">MOSS-Image-Base-8-steps-dmd</a> at strength 1.0</li></ul>Both configs<ul><li>This LoRA's weight: 0.6 – 1.0 — 1.0 is the sweet spot; past ~1.2 colours start to over-saturate</li><li>Negative prompt: <code>worst quality, bad anatomy, watermark, jpeg artifact</code></li><li>Resolution: 1–2.5 MP works well. 16-pixel multiples only.</li></ul><h3 id="aspect-ratio">Aspect ratio</h3>Pick by subject, the way you would for any real picture:<ul><li>Portrait / character work → vertical, 3:4 (e.g. 1376×1840)</li><li>Environment / story / wide scene → horizontal, 16:9 (e.g. 2048×1216)</li></ul><h3 id="no-trigger-word">No trigger word</h3>This is an unconditional style LoRA. There is no trigger phrase — just load it and adjust the weight. The style becomes the default rendering.<h3 id="what-it-is-good-at">What it is good at</h3><ul><li>Cinematic character portraits with dramatic lighting</li><li>Poster-like graphic compositions and bold typography</li><li>Story beats: confrontation, tension, quiet moments</li><li>Environments and wide establishing shots</li><li>Chinese text rendering — MOSS-Image handles CJK typography unusually well, so Chinese poster work is a genuine strength here</li></ul><h3 id="how-it-was-trained">How it was trained</h3><ul><li>168 images, captioned in plain natural language (content only, no style words — so the style is learned as the model's default, not as a prompt trigger)</li><li>MOSS-Image Base, rank 16 / alpha 16</li><li>50 epochs, learning rate 5e-5, AdamW, EMA off</li><li>Fixed clean-end share 40% to bias training toward layout and composition</li><li>Trained with the Fizgig trainer</li></ul><h3 id="sample-metadata">Sample metadata</h3>Every image uploaded here carries its full ComfyUI workflow and Civitai-compatible generation parameters embedded in the PNG. Hit Remix on any of them to load the exact graph that made it.<hr /><h2>中文</h2><h3>这是什么</h3>Arcane Style 是一个用于 MOSS-Image 的风格 LoRA，用来复刻《双城之战》那种厚涂油画质感、硬边色块、高饱和紫／青／琥珀色调，以及电影海报式的构图。它是风格 LoRA，不是人物 LoRA。它不锁定某张脸——你可以用它画任何主体，它会用这套绘画语言去表现。<h3 id="base">重要：请在 Base 模型上使用</h3>这个 LoRA 是用 MOSS-Image Base 训练的，在 Base 上的表现明显好于 SFT。SFT 经过美学后训练，会稀释 LoRA 的风格信号。Base 训练，Base 使用。<h3>两种跑法 —— 我更推荐快的那种</h3>有两种可用的配置。效果都不错，第二种快得多，而画质损失很小。① 默认 —— 50 步，不加速。 最直接的用法，不需要额外下载任何东西。② 推荐 —— 8 步 + DMD 加速 LoRA。 搭配 Base 的 8 步蒸馏版：<a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd">lodestones/MOSS-Image-Base-8-steps-dmd</a>它与本 LoRA 可以干净地叠加。2.5 MP 的图约 12 秒一张，画质只有很小的差异——这里所有样图都是这么出的。<h3>推荐参数</h3>配置 ① —— 默认（50 步）<ul><li>底模： MOSS-Image Base（不要用 SFT）</li><li>采样器： euler</li><li>调度器： simple</li><li>步数： 50</li><li>CFG： 3.0 – 4.0</li><li>Shift： 5.0</li></ul>配置 ② —— 推荐（8 步 + 加速 LoRA）<ul><li>底模： MOSS-Image Base（不要用 SFT）</li><li>采样器： euler</li><li>调度器： simple</li><li>步数： 8</li><li>CFG： 1.0</li><li>Shift： 5.0</li><li>加速 LoRA： <a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd">MOSS-Image-Base-8-steps-dmd</a>，权重 1.0</li></ul>两种配置通用<ul><li>本 LoRA 权重： 0.6 – 1.0，1.0 是甜点；超过约 1.2 色彩会开始过饱和</li><li>负向提示词： <code>worst quality, bad anatomy, watermark, jpeg artifact</code></li><li>分辨率： 1–2.5 MP 都不错，边长必须是 16 的倍数</li></ul><h3>画幅建议</h3>按题材选择，就像真实拍摄那样：<ul><li>人像／角色 → 竖幅 3:4（如 1376×1840）</li><li>环境／故事／宽景 → 横幅 16:9（如 2048×1216）</li></ul><h3>没有触发词</h3>这是一个无条件风格 LoRA。没有触发词——加载后直接调权重即可，风格就是默认渲染。<h3>它擅长什么</h3><ul><li>电影感的角色肖像与戏剧性打光</li><li>海报式的图形化构图与粗体排版</li><li>故事性瞬间：对峙、紧张、静默的时刻</li><li>环境与宽阔的场景建立镜头</li><li>中文字体渲染——MOSS-Image 对中日韩文字的处理异常出色，中文海报是这个 LoRA 的一个真实优势</li></ul><h3>训练方式</h3><ul><li>168 张图，用自然语言打标（只描述内容、不含风格词——这样风格会被学成模型的默认渲染，而不是需要提示词触发的开关）</li><li>MOSS-Image Base，rank 16 / alpha 16</li><li>50 个 epoch，学习率 5e-5，AdamW，关闭 EMA</li><li>Clean-end share 固定 40%，让训练偏向布局与构图</li><li>使用 Fizgig 训练器</li></ul><h3>样图信息</h3>这里上传的每张图都在 PNG 中内嵌了完整的 ComfyUI 工作流和 Civitai 兼容的生成参数。任意一张点 Remix 就能载入生成它的完整节点图。

Files with SHA256: a3ab83000465beee60ac11aa5785dab414fbda82b780303c6946fe0d52fb2613