文件清单 线路：镜像（hf-mirror.com） · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
MOSS_Semi-realistic_Render
作者：baowy
MOSS_Semi-realistic_Render
An experimental strong-style LoRA for MOSS-Image — recommended weight 1.0–2.0, with much heavier stylization at 1.5–2.0. ---
What it does
This is a style LoRA, not a character or object LoRA. It pushes MOSS-Image toward a semi-realistic rendered look: stylized materials, stronger shape language, and a more illustrative sense of light than the base model produces on its own. It was trained on 534 frames pulled from Chinese 3D animation and fantasy short films, with captions written as plain content descriptions (subjects, actions, settings) and no style words — so the style itself lives entirely in the LoRA weights rather than in a trigger token. There is no trigger word. Just load it and set your weight. ---
Recommended settings
| Parameter | Value |
|---|---|
| Weight | 1.0 – 2.0 |
| Sweet spot | 1.0 – 1.2 |
| Strong stylization | 1.5 – 2.0 |
| Base model | MOSS-Image (Base / SFT) |
| Sampler | er_sde |
| Steps | 8 (with DMD accelerator) / 10–20 without |
| CFG | 1.0 |
| Resolution | 1056×1408 or any MOSS-native size |
Best used together with the 8-step DMD accelerator LoRA at strength 1.0 for fast iteration. ---
Weight guide
- 0.5 — very subtle. Barely a style change; useful when you only want a faint hint.
- 1.0 — balanced. Style is clearly present while the base prompt still controls composition.
- 1.5 — strong. The render look dominates; details get more stylized and edges harder.
- 2.0 — maximum. Very heavy stylization, and the model starts to override prompt details — expect composition and subject changes. Recommendation: start at 1.0, push to 1.5–2.0 when you want the style to take over. ---
Notes and limitations
This is an experimental release and it is not a clean, narrow style filter. A few things worth knowing before you use it:
- Human subjects skew female. The training data was human-heavy, which means at high weights a male subject may drift toward a feminine appearance. Keep male prompts at 1.0 or below.
- Abstract / architectural prompts get reinterpreted. At 1.5+ the model may add elements that were not asked for (waterfalls, extra structures, decorative figures) in order to satisfy the style.
- Creature and beast prompts can absorb people. At 2.0 you may find figures appearing in scenes where you asked for none.
- No trigger word to memorize — but also no way to switch the style off other than lowering the weight. If you want a controlled, predictable result, stay at 1.0–1.2. If you want the style to fight your prompt and win, go 1.5–2.0. ---
Training summary
| Base | MOSS-Image |
| Dataset | 534 frames, Chinese 3D animation / fantasy shorts |
| Caption style | content-only, no style tokens, no trigger word |
| Rank / Alpha | 64 / 64 |
| Recommended use | style transfer, mood pieces, stylized illustration |
中文说明
MOSS_Semi-realistic_Render 是一个实验性的强风格 LoRA，推荐权重 1–2；在 1.5–2 时风格化会明显加强。 这是一个风格 LoRA，不是角色或物体 LoRA。它把 MOSS-Image 推向半写实渲染的观感：材质更有造型感、形体语言更概括、光线更接近插画的处理方式，而不是基础模型自带的那种直白写实。 训练素材是 534 帧中国风 3D 动画与奇幻短片，标注全部写成纯内容描述（主体、动作、场景），不写任何风格词，因此风格完全藏在权重里，没有触发词，挂上直接调权重就行。
推荐参数
| 项 | 值 |
|---|---|
| 权重 | 1.0 – 2.0 |
| 甜点 | 1.0 – 1.2 |
| 强风格化 | 1.5 – 2.0 |
| 底模 | MOSS-Image (Base / SFT) |
| 采样器 | er_sde |
| 步数 | 8（配 DMD 加速）/ 10–20（不配） |
| CFG | 1.0 |
| 分辨率 | 1056×1408 或 MOSS 原生任意尺寸 |
建议配合 8 步 DMD 加速 LoRA（强度 1.0）一起用，出图快、便于反复调。
权重感觉
- 0.5 — 非常轻微，只有一点点风格倾向
- 1.0 — 平衡，风格明确但构图仍由提示词主导
- 1.5 — 强，渲染感压过提示词，细节更概括、边缘更硬
- 2.0 — 极限，风格化很重，但模型开始改写你的提示内容
需要注意
这是实验版，不是干净克制的滤镜，有几件事提前说清楚：
- 人物偏女性：训练集人物占比高，高权重下男性主体可能被推向女性外观，男性提示建议压在 1.0 以内
- 建筑 / 抽象场景会被再解读：1.5 以上模型可能自行添加瀑布、额外结构、装饰人物
- 巨兽类提示会混入人：2.0 时可能出现你没要求的人物 想要可控结果就停在 1.0–1.2；想让风格盖过提示词就上 1.5–2.0。 --- Feedback and test images welcome — this is an early experiment and the next version will address the subject-balance issues above.