ShenDuMao-MOSS-Edit-LoRA-v2

神都猫 MOSS 单图编辑 LoRA v2 / ShenDuMao MOSS Edit LoRA v2

神都猫使用 AI-Toolkit 训练，2000 步，v2。** 这是单参考图编辑 LoRA，本页示例使用 **MOSS Image SFT** 推理。

Trained by ShenDuMao (神都猫) with AI-Toolkit, 2,000 steps, v2.** A single-reference edit LoRA; examples here were generated with **MOSS Image SFT**.

插件与下载 / Plugin and downloads

• *ComfyUI 插件、安装教程、中英文技术介绍、工作流 / Plugin, installation, bilingual technical notes and workflows:** ComfyUI-ShenDuMao-MOSS-Edit
• *LoRA 权重 / LoRA weights:** moss wendingedit lr5e-5 1024 v2 000002000.safetensors
• *模型首页 / Model page:** ModelScope

安装和使用 / Installation and usage

• 更新到最新版 ComfyUI；在 ComfyUI/custom nodes/ 执行下方命令并重启。

Update ComfyUI, clone the plugin into custom nodes , and restart.

• 下载本页 LoRA，放到 ComfyUI/models/loras/ 。MOSS SFT 放在 models/diffusion models/ ，完整 Qwen3.5 视觉/文本编码器放在 models/text encoders/ ，FLUX.2 VAE 放在 models/vae/ 。

Put this LoRA in ComfyUI/models/loras/ . Place MOSS SFT in models/diffusion models/ , the full Qwen3.5 vision/text encoder in models/text encoders/ , and the FLUX.2 VAE in models/vae/ .

• 导入 SFT + v2 工作流，选择模型和 LoRA，上传参考图并输入完整编辑指令。编码器必须包含视觉权重，不能是 text-only 版本。

Import the workflow, select the model and LoRA, upload one reference, and enter the full instruction. A full vision encoder is required.

• 起始参数 / Starting settings: **50 steps, CFG 3, Euler, shift 5, LoRA 1.0, seed 42**, empty negative prompt. 保持原图比例，宽高为 16 的倍数。Use the reference aspect ratio and dimensions divisible by 16.

• 使用工作流自带的 sigma 节点；Prepare 内已处理负面分支第 22 层跳过，无需再加 SkipLayerGuidance。Keep the supplied sigma node; negative-branch layer-22 skipping is already handled internally.

提示词 / Prompt format

写清楚数量、位置、原来的外观、目标外观和保留项。Specify counts, locations, existing appearance, target appearance and what should remain.

图片案例 / Examples

原图均由 MOSS SFT 生成；编辑结果直接来自 SFT + 本 LoRA。未通过修图、遮罩合成或其他模型修复，PNG 保留实际工作流。References were generated with MOSS SFT. Edited images are direct SFT + v2 outputs without retouching, masks, compositing or another model's repair. PNGs contain the actual workflows.

添加金框眼镜 / Add gold-rimmed glasses

眼镜添加生效，人物与背景整体保留，脸部和纹理有轻微重绘。The glasses were added with broadly preserved identity and setting; facial and texture details were slightly redrawn.

皮包包身改色 / Handbag body recolor

包身正面、翻盖、侧面由焦糖棕变雾蓝；金属扣与走线大体保留。提手仍棕色，并未完成所有皮革部位改色。The front, flap and sides changed from caramel brown to powder blue. The clasp and stitching were broadly preserved; the handle remained brown, so not every requested leather part changed.

更多类型与指令复测 / More image types and explicit instructions

写真、家居、美食和水彩插画的实际结果如下，同时包含有效修改和未完成项。对照图只做缩放、拼排与加说明，没有修复生成内容。The following selected comparisons include effective edits and remaining limitations; only layout and labels were added.

皮包：明确部件后，改色范围更完整 / Handbag: explicit part descriptions improve coverage

从只改翻盖和侧面，到正面包身也变蓝；金属扣与走线大体保留，提手仍为棕色。 Explicit part descriptions extended the recolor to the front body. The handle still remained brown.

写真：白色针织毛衣改蓝 / Fashion photo: recoloring a knitted sweater

毛衣改蓝生效，街景与姿势大体保持；脸部、衣领和针织纹理仍有重绘。 The sweater became blue; pose and street layout broadly stayed, while face, collar and knit details changed.

家居：沙发改色与抱枕保留的边界 / Interior: sofa recolor and cushion preservation limits

沙发变绿；详细指令改善了左侧黄抱枕，但右侧保留不足，并出现额外绿色物体。 Sofa recolor worked. Yellow-cushion preservation remained incomplete, with an unwanted green object.

美食：明确六颗草莓后，仍未完整替换 / Food: six positions targeted, incomplete fruit replacement

六处位置均得到处理，但仍有红色草莓果肉和绿叶，呈混合形态，不算完整替换成功。 All six positions were targeted, but strawberry flesh and leaves remained: object replacement was incomplete.

水彩插画：斗篷部分改色，背景整体保留 / Watercolor: partial cape recolor and broad scene preservation

前方衣片变红，外侧仍有浅蓝；详细指令使尾巴保留橘色，但轮廓和姿态仍改变。 Front cape panels became red; blue fabric remained outside. Orange tail color improved, but its shape changed.

工作流分区 / Workflow layout

下载更多对照图库 / Download the comparison gallery

技术依据 / Technical basis

配套插件参考 Ostris AI-Toolkit 的 MOSS 编辑推理约定，实现两路图像条件、隔离参考注意力、采样期 K/V 缓存和对齐的 sigma 调度。Plugin details and upstream attribution are documented in the GitHub repository.

当前为实验性编辑权重；局部细节可能重绘，数量、物体替换、背景更换等指令不保证完整完成。This is experimental: non-target details may be redrawn, and counting, object replacement or background edits may not be completed.

署名与许可 / Credits and license

© 2026 神都猫 / ShenDuMao. Trained with AI-Toolkit, 2,000 steps, v2.**

LoRA: Apache License 2.0. 配套插件代码为 MIT，并保留 Ostris 的原始署名。Plugin code is MIT with upstream attribution retained. MOSS 底模与其他组件分别遵循各自许可。Base models and other components retain their respective licenses.