{
  "$type": "site.standard.document",
  "canonicalUrl": "https://handyman5.github.io//posts/making-a-blog-logo",
  "path": "/posts/making-a-blog-logo",
  "publishedAt": "2024-07-11T17:30:27.000Z",
  "site": "at://did:plc:yxhvd47p53gmb5zfiktqq3og/site.standard.publication/3mohn6mqvab25",
  "tags": "meta",
  "textContent": "To commemorate the new blog, I wanted to have a cool logo up in the corner. Since one of my current interests is generative AI, I thought I'd try using it to generate myself a logo. This post explains the process.\n\n\n\nGetting Started\n\nI already have the kit set up to run a large language model (LLM) in my home network. However, I've only ever used it for text generation before; I had no idea how to make it generate images.\n\nI started with Koboldcpp, since I recalled that it was able to generate images as of a recent release. This wiki page explained how to generate images with Koboldcpp, and provided some example models to try. I poked at a couple of different ones before I tried the Anything v3 model linked from that page. I ran the model with --sdmodel <path-to-anything>.\n\nI admit, it took me a while to figure out how to even start trying to generate an image! I had to poke around in the text generation UI (\"kobold lite\") for a while before I realized that the image generation UI was an entirely different path: <IP>:5001/sdui.\n\nPrompting the Model\n\nOnce I had that running, I had to figure out how to actually prompt the model and get images out. A useful technique I ran across is to use an empty white square with the img2img option to force the model to generate something that's mostly white, which is what I wanted for my logo. (I used the convert command to generate said image.)\n\nI wound up with:\n\nThe settings I used were:\n\n Prompt: sparse black line drawing vector logo icon, stylized male human head filled with gears and circuits, profile, no text, no fill, line art, adobe illustrator, vector art\n Negative prompt: color, blurry, complex, intricate, neon, mecha, robotech, spiky, anime\n Sampler: Euler a\n Steps: 40\n* Guidance: 15\n\nTakeaways\n\nThis took a lot of work! Both to get the image generation pipeline working in the first place, and also to tweak the parameters and try to get to an image I liked. Even this one isn't exactly where I'd want it, but I think it's about as good as it's going to get for my current level of setting-up-the-pipeline and prompting skills, so I'm setting it down for now. I did appreciate the ability to translate the rough image I had in mind to something tangible; I have no drawing talent whatsoever so it was nice to have borrowed access to this skill.",
  "title": "Making a Blog Logo"
}