AI 生成画像のテキスト レンダリング

Writing is an unforgiving target. A face can be slightly off and still read as a face, but one wrong stroke turns an 'e' into a 'c'. Early diffusion models learned that shop signs, book covers and T-shirts contain lettering, so they produced shapes with the rhythm of text while getting the actual letters wrong. The text encoder was a major cause. Stable Diffusion 1.x used OpenAI's CLIP text encoder, which was trained to match whole captions to images. CLIP splits words into subword tokens, so a word like 'bakery' may become a single token, and the image model receives a representation of meaning rather than a sequence of letters. Nothing in that signal says 'b, then a, then k'. Google's Imagen research in 2022 found that a large frozen T5 language-model encoder improved prompt following, and its samples rendered text noticeably better. A follow-up Google study on character-aware models showed that byte-level encoders such as ByT5, which see individual characters, spell much better. Resolution mattered too. Small letters occupy very few cells in a compressed latent, so fine strokes blur. Training data was another limit. Captions rarely wrote out the text visible in an image, so the model could not link words to letter shapes. Progress came from several directions. DeepFloyd IF and Stable Diffusion 3 paired their image models with T5 encoders. Research systems such as TextDiffuser and GlyphControl gave the model explicit layouts or pre-rendered letter images to follow. Products including Ideogram, DALL-E 3 and later autoregressive image generators made legible text routine, helped by training captions that include the words shown in the image. One misconception is that modern models read letters from a font file. They still generate letters as pixels, which is why long passages, rare words and small print still go wrong.