rw-book-cover

Metadata

Highlights

  • Yes, it’s indeed one of the world’s most advanced large language models. Many people use it for chatting, consulting, and document reading. With the introduction of GPT-4’s capability to process web content, PDFs, and images, it opens up a wide range of possibilities, including generating prompts for uploaded images. (View Highlight)
  • Not bad, but we’re still a bit away from our goal. That’s okay; let’s work on refining it. Here, I described the text structure I want for the prompt (personal preferences may vary), and I also hope it uses concise, clear, and objective language for description. (View Highlight)
  • It’s getting better, and it’s already better than most labeling tools, isn’t it? But experienced trainers won’t stop here. I asked it to refer to the well-known tags from the Danbooru/Gelbooru specific dataset and to prioritize using these well-known tags in labeling images in similar prompts. (View Highlight)
  • Great, this already goes far beyond my own ability to craft prompts. Of course, let’s add a bit more complexity. My personal preference is to describe the visual style and theme of the scene in the beginning of the text. Now, let’s have it do that. (View Highlight)
  • Perfect! At least for me, I’m really starting to like this GPT guy more and more! Now, I’ll have it summarize all the rules mentioned above and put together a set of text for future use. (View Highlight)
  • While writing this article, I learned about the emergence of GPTs and there will be more GPT-based apps in the future, which is simply fantastic, isn’t it? I’m also trying to use it to create my own prompt robot. Right now, it’s still a work in progress, and it doesn’t always cooperate as expected. I’m constantly fine-tuning it. When the time is right, I think I might share it with others to benefit more people. Well, that’s the sharing for today. I have to admit that there were many challenges in debugging GPT. I simplified some of the steps here, but the overall logic remains unchanged. Once you successfully debug it, congratulations, you’ll have an unprecedented, smart, and convenient labeling tool at your disposal (View Highlight)