🐼 stable diffusion角色设计 用Controlnet分别在SD WEBUI和ComfyUI中绘制多角度人像的方法

The video begins with an introduction to the 28th episode of the SD tutorial series, focusing on character design. The host discusses the widespread interest in AI painting since its inception and shares their experience with the topic. They mention a recent article by an international expert on using SD for character design, which they found useful and decided to share in a tutorial. The host emphasizes the importance of consistency in character design across different angles, traditionally achieved through 3D modeling. However, the article introduces a novel method using SD and Lora training, specifically for characters created entirely through AI-generated images. The process involves creating a sequence of multi-angle character images, training Lora with these images, and refining the results to achieve a consistent character design. The host also provides a link to the article in the video description and outlines the three main stages of the process: creating a sequence of character images, training Lora, and refining the results.

This paragraph delves into the specifics of Lora training and the challenges associated with it. The host explains that Lora training typically relies on real-life人物 or existing素材, which may not align with AI-generated characters. The article's author addresses this by proposing a method to train Lora using 100% virtual characters. The process is divided into three stages: generating a sequence of multi-angle character images, training Lora with these images, and refining the model. The host emphasizes the importance of the image sequence, as it allows for image segmentation and Lora training. They also discuss the practical aspects of training, such as the minimum number of images required and the recommended image size for Lora training. The host shares their实操 experience, including the use of腾讯云16G显存 for image generation and the potential issues with lower VRAM.

The host transitions to a practical demonstration of character design using SD. They guide the audience through the process of using the templates provided in the article, including a 15-angle facial template and a 15-grid image for SD's understanding of individual images. The host explains the use of CONTROLNET for image segmentation and the selection of open pose and lineart models for facial direction and detail control. They also discuss the use of tile models and upscaling algorithms to enhance image quality. The host demonstrates the generation of a simple image using the 'one girl' prompt and shows how to放大 the image using tile models and upscaling algorithms. They then provide a detailed explanation of the CONTROLNET models' roles and how they contribute to the final image quality.

The host introduces the deployment of the character design system using COMFY UI, which allows for a more streamlined and batch processing of images. They explain the system's capability to output multiple sets of images and discuss the use of CONTROLNET models for image control. The host also shares their拓展用法 of the 15-grid template to a 4-grid template and the rationale behind using two CONTROLNET open pose models to avoid facial duplication in the grid. They provide insights into the impact of prompt words on the final image, emphasizing the importance of consistent seed values and sampling steps. The host also discusses the use of different models and their impact on image quality, offering suggestions for optimizing the training process and addressing common issues such as facial recognition inaccuracies and image degradation.