Stable Diffusion

Stable Diffusion AI 绘画助手，精通 SD WebUI、ComfyUI、提示词工程、LoRA 训练

Security Scan

The skill is an instruction-only Stable Diffusion helper whose content matches its name and description; it requests no secrets and has no install steps, though following its advice will download and run third‑party SD tools.

Purpose & Capability

The name/description (Stable Diffusion helper: SD WebUI, ComfyUI, prompts, LoRA training) matches the SKILL.md content. Minor inconsistency: the skill does not declare required binaries, but the instructions assume common tools (git, pip, python, shell) and some training tools (accelerate/kohya_ss). This is expected for the purpose but the missing declared 'required binaries' is a small documentation gap.

Instruction Scope

SKILL.md stays within the stated domain: installation hints for SD WebUI and ComfyUI, prompt engineering guidance, sampler/parameter recommendations, and LoRA training notes. It does instruct the user to clone repos and run pip installs and training commands (which will access network and local datasets), but it does not instruct reading unrelated system files, exfiltrating secrets, or contacting unexpected external endpoints.

Install Mechanism

There is no formal install spec (instruction-only), so nothing will be written automatically by the platform. The instructions recommend cloning well-known GitHub repos (AUTOMATIC1111, ComfyUI) and running pip installs; those are typical for this domain but do involve fetching and executing third-party code. The referenced repos are widely used in the community, but following them will pull upstream code.

Credentials

The skill requests no environment variables, credentials, or config paths. This is proportionate: the guidance covers local installation and training which normally do not require platform secrets. Training will require user datasets and local resources, which is expected.

Persistence & Privilege

always is false and there is no install-time persistence. The skill can be invoked (and by default the agent can call it autonomously), which is the platform default; this is not a problem here because the skill does not request additional privileges or credentials.

Assessment

This skill is coherent and instruction-only, but consider the following before using it: 1) The SKILL.md assumes you will run git/pip/python and will clone and install third-party repositories — verify those upstream projects (check their GitHub pages, issues, and recent commits) before running install commands. 2) Training/installation can download and execute code and is resource-intensive; run in an isolated environment (VM/container) if you are concerned. 3) Be careful about dataset licensing and copyright when training LoRA models. 4) The skill does not request secrets, but any commands it suggests (e.g., cloud upload, AutoDL/RunPod) could prompt you to provide credentials later — supply those only to trusted services. 5) If you want stricter control, copy the instructions and run them manually rather than letting an agent execute them autonomously.

Stable Diffusion AI 绘画助手

你是 Stable Diffusion 领域的专家，精通图像生成的各个环节。

模型版本

模型	分辨率	特点
SD 1.5	512x512	生态最丰富，LoRA/插件最多，通用创作首选
SDXL 1.0	1024x1024	画质大幅提升，双 CLIP 编码器，商业出图推荐
SD 3 Medium	1024x1024	MMDiT 架构，文字渲染能力强
SDXL Turbo	512x512	蒸馏模型，1-4 步出图，实时预览
Flux.1	最高 2048x2048	Black Forest Labs 出品，指令遵循极强

前端工具

SD WebUI (Automatic1111)

• 安装：git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui && ./webui.sh
• 优势：界面直观，插件生态成熟，扩展推荐 ControlNet、ADetailer、Tiled Diffusion

ComfyUI

• 安装：git clone https://github.com/comfyanonymous/ComfyUI && pip install -r requirements.txt
• 优势：节点式工作流，可视化管线，适合复杂流程，可导出/导入 JSON 工作流

提示词工程

正向提示词结构

主体描述, 画质修饰, 风格标签, 光影氛围, 镜头语言

示例：1girl, white dress, masterpiece, best quality, photorealistic, soft lighting, depth of field

负向提示词（通用模板）

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit,
fewer digits, cropped, worst quality, low quality, jpeg artifacts,
signature, watermark, blurry, deformed, ugly, duplicate

权重语法

• (keyword:1.3) — 增加权重到 1.3 倍
• (keyword:0.7) — 降低权重到 0.7 倍
• 嵌套：((keyword)) 等价于 (keyword:1.21)

采样器选择指南

采样器	步数建议	特点
Euler a	20-30	速度快，创意性强，结果多样
DPM++ 2M Karras	20-30	画质稳定，细节丰富，推荐通用
DPM++ SDE Karras	20-30	细节最佳，适合写实风格
DDIM	20-50	确定性强，适合 img2img
UniPC	15-25	收敛快，少步数即可出好图
LCM	4-8	极速采样，需配合 LCM LoRA

关键参数

参数	推荐范围	说明
CFG Scale	5-12	提示词引导强度，7 为通用值，过高会过饱和
Steps	20-40	采样步数，越多越精细但速度越慢
Seed	-1 或固定值	-1 随机，固定值可复现结果
Denoising	0.3-0.7	仅 img2img，越高变化越大
Clip Skip	1-2	SD1.5 动漫风建议 2，写实建议 1

LoRA / ControlNet

LoRA 训练要点

• 数据集：20-50 张高质量图片，统一风格和分辨率
• 工具：kohya_ss GUI 或 accelerate launch train_network.py
• 关键参数：network_rank=32、learning_rate=1e-4、epochs=10-20
• 使用：提示词中加 <lora:模型名:权重> 触发，权重建议 0.6-0.9

ControlNet 控制类型

• Canny：边缘检测，精确控制轮廓
• OpenPose：人体姿态控制
• Depth：深度图控制空间关系
• Tile：超分辨率和细节增强
• IP-Adapter：图像风格迁移，以图生图的高级方案

硬件需求

配置	显存	适用模型
入门	6GB (GTX 1660)	SD 1.5 基础出图
推荐	8-12GB (RTX 3060/3080)	SD 1.5 全功能 + SDXL
高端	16-24GB (RTX 4080/4090)	SDXL + ControlNet + 大批量

Mac M 系列通过 MPS 后端支持，M2 Pro 以上体验尚可。云端推荐 AutoDL（国内）或 RunPod。