3 Steps to Automate Original Illustration Generation with Stable Diffusion [Beginner's Guide]
機械翻訳 / Machine-translated
A
AI Friends
4日前・12分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What you'll learn in this article
Stable Diffusion is an AI tool that automatically creates illustrations and images simply by typing a text description. As of August 2026, the latest version — Stable Diffusion 3.5 — has been released, and its biggest appeal is that it can be used for free, as many times as you like. Other image-generation AIs limit you to a few dozen images per month or charge a monthly fee, but Stable Diffusion is open-source (a system anyone can use freely), so there are no usage limits. In addition, the images you generate can be used commercially, meaning you can use them directly for business purposes such as social media post illustrations, YouTube thumbnails, and blog graphics. Because it can be installed on your own computer without any programming knowledge, you can also create sensitive images with peace of mind.
First, set up your environment for using Stable Diffusion. For beginners, the recommended approach is to install a free application called "Stable Diffusion web UI" on your computer. What you need is a Windows or Mac computer, an NVIDIA graphics card (RTX 3060 or higher is recommended), and at least 16 GB of RAM. Even if your computer does not have a dedicated graphics card, you can run it in a browser alone by using Google Colab (a free cloud environment provided by Google). The installation steps are simple: download the latest version from the official GitHub page, extract the folder, and double-click the executable file inside. On first launch, the necessary files are downloaded automatically, so wait about 10 minutes. Once startup is complete, the control interface opens automatically in your browser.
Once the control interface opens, configure the settings for generating illustrations. At the top of the screen is an input field called "Prompt," where you type in English what kind of image you want drawn. For example, writing something specific like "a cute cat sitting on a chair, anime style" will get you closer to the image you have in mind. In the "Negative Prompt" field below that, write the opposite — things you do not want in the image. For example, entering "low quality, blurry" helps avoid poor-quality results. Next, set the "Sampling Steps" on the right side of the screen from 20 to 30. A higher number produces a more carefully rendered image, but it also takes more time. You can leave "CFG Scale" at 7. This is the value that determines how faithfully the prompt is reproduced.
Once configuration is done, click the orange "Generate" button at the bottom of the screen. The image will then appear in real time on the right side of the screen. Depending on your computer's specs, it takes roughly 30 seconds to one minute to complete. Generated images are automatically saved to the folder outputs/txt2img-images inside the web UI directory. If the result is not what you expected, simply rewrite the prompt and press "Generate" again — you can redo it as many times as you like. Even with the same prompt, changing the "Seed" (a random number value) produces a different image each time. When you get an image you like, copy the settings information displayed below it so you can reproduce a similar image later. The finished image is saved in PNG format, so you can upload it directly to social media or edit it further with image-editing software.
There are three common pitfalls for beginners. The first is the phenomenon of "only a completely black image appearing." This is caused by insufficient graphics card memory, and can be resolved by reducing the image size to 512×512 or adding the startup option --medvram to the configuration file. The second is the problem of "prompts written in Japanese not working as expected." Because Stable Diffusion was trained primarily on English, accuracy drops with Japanese input. Convert your text to English using DeepL or Google Translate before entering it. The third is the failure where "faces and hands come out distorted and messy." To prevent this, add "bad anatomy, extra fingers" to your negative prompt, or check the "Restore faces" option to improve results. Changing the model data can also significantly affect image quality, so downloading a highly rated model from a model distribution site such as Civitai and giving it a try is also recommended.
Once you are comfortable with the basics, you can try more advanced usage. The first is using additional training data called "LoRA." By incorporating LoRA, you can easily reproduce a specific character or art style. For example, LoRA files that mimic the art styles of famous studios such as "Studio Ghibli style" or "Pixar style" are distributed for free. The second is the "img2img (image-to-image)" feature. By uploading a rough sketch or photo you have created yourself, the AI will use it as a base to finish the artwork. You can transform a rough sketch into a beautifully rendered illustration, or turn a daytime photo into a nighttime scene. The third is an extension called "ControlNet," which lets you specify poses and compositions in fine detail. Simply by drawing a stick figure line drawing, the AI will render a character in that exact pose, making it dramatically easier to decide on illustration compositions. By combining these features, you can produce illustrations of professional-grade quality.
This article is a cross-post from AI Friends.
