超详细！AI 绘画神器 Stable Diffusion 基础教程

This section focuses on ControlNet, a crucial plugin for Stable Diffusion that enables precise control over image generation. Before ControlNet, AI art generation was often unpredictable. ControlNet elevates AI art to a productivity level by allowing detailed manipulation of generated images. The article details eight primary ControlNet application models: OpenPose for precise human pose control (including multi-person and hand poses), Canny and HED for edge detection to replicate compositions or color line art, Scribble for higher creative freedom from sketches, Mlsd for architectural line structure detection, Seg for semantic segmentation to alter large scene styles, Normal Map for replicating 3D object lighting, and Depth for maintaining structural depth. It also explains the power of combining multiple ControlNet models for multi-conditional control and how this can be used to generate animations by controlling pose across frames.

This section provides practical guidance on deploying and installing Stable Diffusion, outlining three main methods: cloud deployment, local deployment, and one-click installation. Cloud deployment via Google Colab is recommended for users with less powerful hardware, offering free access to powerful GPU resources for fast image generation, though with time limitations. The article links to recommended Colab projects for Stable Diffusion and ControlNet. Local deployment offers greater extensibility, customization, privacy, and freedom but requires high-end hardware (NVIDIA GPU with 8GB+ VRAM, 16GB+ RAM for Windows; M1/M2 chip for Mac). Step-by-step instructions are provided for local installation on Mac M1, covering Homebrew, Python, cloning the Stable Diffusion WebUI repository, downloading and organizing models (Checkpoint, LoRA, Textual Inversion), and running the WebUI. Instructions for installing the ControlNet extension and troubleshooting ControlNet on Mac M1 are also included. For simplicity, DiffusionBee is presented as a one-click installation option, though with limited extensibility.

This section focuses on the art of prompt engineering, emphasizing its critical role in achieving high-quality image generation with Stable Diffusion. It introduces a structured four-part prompt format: image quality/style, main subject, scene details, and style reference. Key prompt syntax techniques are explained, including word order importance, weighting (e.g., `(word:1.3)`), mixing elements (`|`), using `+` and `AND` for conjunctions, and staged rendering (`[A:B:step]`). The use of positive prompts for quality enhancement (e.g., `masterpiece, best quality`) and negative prompts to exclude unwanted elements (e.g., `nsfw, bad hands`) is also covered. The section further suggests leveraging ChatGPT as an assistant for generating prompt ideas and references, providing a link to a guide on using OpenAI's ChatGPT for Stable Diffusion prompts.

The article concludes by summarizing the capabilities of Stable Diffusion, emphasizing how ControlNet for precise image control and various model loading/training techniques for style customization have transformed AI art into a practical production tool. It highlights that these advancements empower designers to significantly enhance their design efficiency by leveraging these powerful AI tools.