ELI5
Explain it like I’m five日本語English
In one lineAI image generation is artificial intelligence that creates new images from text prompts and other input, based on patterns learned from data.
AI image generation is AI that makes pictures from words. Type something like “a cat swimming in a sunset sea.” It makes a drawing or a photo-like image to match. It learned from lots of pictures paired with descriptions. That is how it knows how words and looks go together. A common method starts with something like TV static. Then it removes the noise bit by bit, until a picture appears. Anyone can make a picture in seconds. But rights to the pictures used for learning are being debated. It is a friend to people who can’t draw. And a headache for people who can.
Most tools today use a “diffusion model.” In training, clear pictures are slowly buried in noise, like TV static, and the AI learns the reverse: how to remove a little noise at a time. To make a new picture, it starts from pure static and removes noise dozens of times, steering toward what your words describe. Many models do this in a squeezed-down version of the image, which is why some can even run on a home computer.
Research on turning text into pictures is older, but in 2022 OpenAI’s DALL·E 2, Google’s Imagen, Stability AI’s Stable Diffusion and Midjourney arrived one after another, and results began to look like real photos and hand-drawn art. Stable Diffusion was released openly, so people around the world built their own versions and add-ons, and it spread very fast.
In Japan, Article 30-4 of the Copyright Act, added in 2018, generally lets works be used without permission for purposes like AI training, where nobody is “enjoying” the work itself. It does not apply when it would unfairly harm the rights holder. And if a finished image looks like an existing work and was based on it, it can infringe, just like a hand-drawn copy. Japan’s Agency for Cultural Affairs published its official view on AI and copyright in March 2024.
Sources:Wikipedia: Text-to-image model、Wikipedia: Diffusion model、Agency for Cultural Affairs: AI and copyright (Japanese)、INTERNET Watch: AI and copyright explained (Japanese)
Have you used an AI image tool?
Tap to see what everyone said.
ChatGPTQ. What is ChatGPT?
Generative AIQ. Does generative AI just copy and paste from the internet?
Large language model (LLM)Q. Does an LLM memorize its answers?
Claude Code
Muse (Meta’s AI)Muse is an AI agent from Meta, the company behind Instagram and Facebook.
Prompt (AI)A prompt is the instruction or question you type into a generative AI.
AI hallucination
Deepfake
Machine learningQ. Do people write all the rules in machine learning?
Deep learning
AI agentQ. Is an AI agent just a chat AI?
GPUA GPU is a part inside a computer that does the math for pictures and video.Source (in Japanese): Wikipediaで「画像生成AI」を調べる