<h1 id="moss-image-base-(5b)-bastyle-lora-(v1.0)">MOSS-Image-Base (5B) - BAStyle LoRA (v1.0)</h1>A BF16 Blue Archive style clean cel-shaded illustration LoRA trained on <a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base">OpenMOSS / MOSS-Image-Base (5B)</a>.<a target="_blank" rel="ugc" href="https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base">OpenMOSS / MOSS-Image-Base (5B)</a> 的 BF16 蔚蓝档案风格清透赛璐璐画风 LoRA。<h2 id="overview">Overview / 简介</h2><ul><li>Style &amp; Scene Coverage: Covers both close-up Memorial Lobby portraits and even-lit story cutscene CGs (multi-character group scenes, lone character overlooking a cityscape, campus architecture, and daytime/nighttime skies). 除近景纪念大厅插画外，同样支持清爽平光的剧情过场 CG（多人日常同框、单人背影远眺）以及学院建筑、昼夜天空等背景空景。</li><li>Feature Decoupling &amp; Prompting Tips:</li><li>Halos: Not baked into the trigger word. Describe halo color and shape when desired; omit halo terms for halo-free characters. 光环按需控制：需要光环时描述颜色与形状；不写光环词汇时默认不出光环。</li><li>Pure Empty Backgrounds / Skies: When generating scenes with <code>no people in the frame</code>, include <code>2D anime background illustration</code> in the opening clause (instead of photographic <code>shot</code>) so the base model does not fall back to realistic photography. Add <code>with faint concentric rings</code> to summon the signature ringed sky. 画面无人物的纯建筑或天空空景建议在开头加入 <code>2D anime background illustration</code> 以避开底模写实摄影先验；描写天空时加入 <code>with faint concentric rings</code> 可生成原作标志性的同心圆环天空。</li><li>Frozen Text Encoder: Trained on the DiT only (<code>Rank 32 / Alpha 16</code>, <code>BF16</code>) with <code>Qwen3.5-9B</code> frozen, preserving native natural-language prompt adherence and in-image text rendering. 冻结 <code>Qwen3.5-9B</code> 仅训练 DiT，保留底模原生的中英文自然语言理解与文字生成能力。</li><li>Dataset: 58 high-resolution illustrations captioned with high-quality natural-language descriptions via Gemini 4 Carbon, with manual verification of edge cases (in-image text and halo annotations). 58 张高清插画采用 Gemini 4 Carbon 高质量自然语言打标，并人工核实了部分细节问题（重点排查了画面内文字与光环标注）。</li></ul><h2 id="comfyui-settings">ComfyUI Settings / 推理参数</h2>A dependency-free native ComfyUI workflow is attached as a Workflow <code>.zip</code> file (and embedded directly inside the sample PNGs).附带无第三方插件依赖的原生 ComfyUI 文生图工作流压缩包（样张原图内同样嵌有完整工作流元数据，下载原图拖入 ComfyUI 即可直接加载）。<ul><li>Base DiT (<code>UNETLoader</code>): <code>moss_image_base_bf16.safetensors</code> or <code>moss_image_base_int8_convrot.safetensors</code></li><li>Text Encoder (<code>CLIPLoader</code>): <code>qwen3.5_9b_int8_convrot.safetensors</code> (or <code>bf16</code>), Type: <code>moss_image</code></li><li>VAE (<code>VAELoader</code>): <code>flux2-vae.safetensors</code></li><li>Latent (<code>EmptyFlux2LatentImage</code>): Multiples of <code>16</code>: <code>1344x896</code>, <code>1536x1024</code>, <code>896x1344</code>, <code>1920x1088</code> (recommended <code>&gt;= 960x640</code>)</li><li>Trigger Word: <code>bastyle,</code> at the start of the prompt</li><li>LoRA Strength: <code>0.90 - 0.95</code> standalone (<code>0.85 - 0.90</code> with character LoRAs)</li><li>Flow Shift (<code>ModelSamplingAuraFlow</code>): <code>5.0</code> (<code>6.0</code> above <code>3 MP</code>)</li><li>Skip-Layer (<code>SkipLayerGuidanceDiTSimple</code>): <code>double_layers = ""</code>, <code>single_layers = "22"</code>, <code>start = 0.0</code>, <code>end = 1.0</code></li><li>Sampler / Scheduler: <code>euler</code> + <code>simple</code>, <code>28 - 32</code> steps, <code>CFG = 3.0</code></li><li>Negative Prompt: Empty (<code>""</code>), or <code>halo, glowing halo above head</code> if strictly suppressing halos</li></ul><h2 id="prompting-guide-and-examples">Prompting Guide &amp; Examples / 提示词写法与完整示例</h2>Do not use Danbooru-style tag lists (such as <code>masterpiece, best quality, 1girl</code>). Write complete natural-language sentences starting with <code>bastyle,</code>.无需堆砌 <code>masterpiece, best quality, 1girl</code> 等传统词条，请在 <code>bastyle,</code> 后直接使用完整的自然语言长句描述画面。<h3 id="prompt-templates">Prompt Templates / 万能句式结构</h3><ul><li>Single Character / Memorial Lobby (单人 / 纪念大厅) <code>bastyle, [Camera angle &amp; shot size] of [hair/eye color, expression, action, optional halo color &amp; shape], wearing [outfit details] in [environment], lit by [light type &amp; direction].</code></li><li>Multi-Character Cutscene CG (多人剧情过场 CG) <code>bastyle, [Wide/medium-wide shot] of [N] students in [scene], lit by clear even daylight from [direction]. Left to right: [Character A details &amp; halo]; [Character B details &amp; halo]; [Character C details &amp; halo].</code></li><li>Empty Architecture / Sky Background (纯无人建筑 / 天空空景) <code>bastyle, A [camera angle] wide 2D anime background illustration of [architecture or sky details + optional "with faint concentric rings"], [lighting], no people in the frame.</code></li></ul><h3 id="verified-sample-prompts">Verified Sample Prompts / 实机验证示例提示词</h3>1. Memorial Lobby - With Halo (纪念大厅 · 带光环)<pre><code>bastyle, An eye-level medium shot of a quiet girl with long dark-violet hair, warm amber-red eyes, and a thin glowing pale-pink geometric halo above her head sitting by a tall sunlit library window, holding a closed hardcover book against her chest and looking calmly toward the viewer. She wears a cream knit cardigan over a white collared shirt and navy pleated skirt, with soft afternoon sunlight streaming through the window from the left.</code></pre>2. Memorial Lobby - Without Halo (纪念大厅 · 无光环)<pre><code>bastyle, An eye-level medium close-up shot of a gentle girl with flowing silver-white hair, luminous sapphire-blue eyes, and a small navy hair ribbon leaning over a sunlit seaside boardwalk railing at golden hour sunset, smiling softly back toward the viewer as the warm sea breeze lifts her hair and red neck ribbon. She wears a crisp cream cardigan over a white school shirt, bathed in translucent amber-gold sunset rim light from the upper left against sparkling ocean waves and pastel orange-and-violet twilight clouds.</code></pre>3. In-Image Text Rendering (画面内英文写字)<pre><code>bastyle, A slight low-angle medium close-up shot of an energetic blonde wolf-eared girl with bright emerald-green eyes, a playful fang smile, and a glowing golden star-ring halo floating above her head standing on a breezy school rooftop under a crisp azure sky with fluffy cumulus clouds, holding up a small wooden cafe chalkboard cleanly written in neat white chalk text "After-School Sweets!" with a cute doodle star beside it, wearing an open pastel-mint hoodie over a white blouse with bright afternoon sunlight from the upper right.</code></pre>4. Story Cutscene CG - Lone Overlook (单人背影远眺城市过场)<pre><code>bastyle, A wide eye-level shot from behind a lone girl with long silver-blue hair in a navy school cardigan standing at a metal observation deck railing in the right foreground, looking out across a sprawling coastal city, harbor bridges, and distant mountain ridges under a calm late-afternoon sky, lit by gentle natural daylight from the upper left.</code></pre>5. Story Cutscene CG - Multi-Character Group (平光多人街头过场)<pre><code>bastyle, An eye-level medium-wide group shot of three students walking abreast across a clean city street crosswalk before low urban buildings and green street trees under a blue sky, lit by clear even daytime sunlight from the upper left. Left to right walk a short-haired blonde student with green eyes and a yellow ring halo waving one hand; a dark-haired student with round glasses and a cyan geometric halo carrying a paper shopping bag in a navy blazer; and a pink-ponytail student with amber eyes and a pale-pink halo holding a canned drink in a beige cardigan and grey plaid skirt.</code></pre>6. World Background - Campus Architecture (学院建筑无人空景)<pre><code>bastyle, A symmetrical eye-level wide 2D anime background illustration across a clean white stone plaza and classical colonnade courtyard leading toward a grand modern-gothic academy clock tower and glass-walled campus buildings under a clear blue sky with faint concentric rings, crisp cel-shaded architectural linework and clean cast shadows across pale stone pavement, no people in the frame.</code></pre>7. World Background - Daytime Ringed Sky (白天同心圆环天空)<pre><code>bastyle, A low-angle wide 2D anime background illustration looking up past a painted school rooftop railing and wire fence into a vibrant clear cyan-blue daytime anime sky filled with painterly white cumulus clouds and giant faint concentric halo rings across the upper sky, bright clean anime daylight, no people in the frame.</code></pre>8. World Background - Nighttime Ringed Sky (黑夜同心圆环星空)<pre><code>bastyle, A wide nighttime low-angle 2D anime background illustration from a quiet hillside road and white guardrail looking up at a deep-indigo starry anime night sky with faint concentric rings, glowing stars, and a crisp crescent moon above stylized dark mountain silhouettes and warm distant town lights in the valley below, cool moonlight, no people in the frame.</code></pre>9. Direct Chinese Prompt Example (纯中文自然语言直出)<pre><code>bastyle, 平视中近景镜头，清晨洒满阳光的学院林荫道上，一位留着深紫色长发、紫晶色眼眸、戴着圆框眼镜的温柔少女抱着一摞精装书回头微笑，身着米色针织开衫、白色衬衫与深蓝色百褶裙，微风轻轻吹起她的发丝，左上角通透的晨光在发梢与肩头勾勒出明亮的高光。</code></pre><h2 id="known-limitations-and-roadmap">Known Limitations &amp; Roadmap / 已知问题与后续计划</h2><ul><li>Facial &amp; Eye Proportions in Standalone Generation: Because the <code>v1.0</code> dataset includes a high proportion of wide story cutscene CGs and multi-character scenes (with relatively few open-eye facial close-ups) and uses high-noise-weighted timestep sampling (<code>logit_mean = -1.09</code>), generating generic original characters standalone may occasionally retain the base model's older anime / classic galgame facial proportions and glossy eye gradients. 由于 <code>v1.0</code> 数据集中远景剧情过场与多人场景占比较高（睁眼正脸特写样本较少），且时间步采样偏向高噪构图区间，单跑生成原创路人角色时，人物脸型比例与眼瞳高光偶尔仍会带有底模原生的早期二次元 / 古早 Galgame 感。</li><li>Current Workaround: For more consistent modern character faces in <code>v1.0</code>, consider pairing this style LoRA (<code>strength = 0.85 - 0.90</code>) with your own character LoRA, or explicitly specifying facial and eye traits in the prompt. 当前建议：若对人物五官画风要求较高，建议自备角色 LoRA 叠加使用（本画风 LoRA 权重设为 <code>0.85 - 0.90</code>），或在提示词中具体描述脸型与眼型细节。</li><li>Planned Fix: Future updates (<code>v2</code> / portrait refinement patch) will expand open-eye close-up portrait samples, refine facial feature annotations, and rebalance low-noise timestep sampling to improve standalone character facial fidelity. 后续计划：未来版本将补充睁眼正脸近景样本、细化五官结构标注并优化中低噪区间采样，修复单跑人物五官偏古早的问题。</li></ul><h2 id="license-and-disclaimer">License &amp; Disclaimer / 协议与免责声明</h2>Released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0). This LoRA is an unofficial fan-made style study intended strictly for non-commercial personal learning, research, and fan creation. Any commercial use, paid generation service, or paid redistribution is prohibited.本模型权重采用 CC BY-NC 4.0（署名-非商业性使用 4.0） 协议发布，仅限个人非商业性质的技术研究、学习交流与同人二创使用，严禁用于任何商业用途、付费代出图或二次售卖。

Files with SHA256: 40cf72663a27ec64d5c6ebbc2baf03e3d5123074a286b1908dfa862a1452a49d