Readable text in AI images: stop asking the model, composite it

SD 1.5 cannot spell. SDXL sometimes manages a short word. Newer models do better, but for a sign, a poster headline or a label you need to be exact, the reliable method is to leave space and set real type afterwards. Why models garble text The model learned what text looks like as texture, not as letters. CLIP tokens split words into pieces, and nothing in the training forces each glyph to be correct. So you get letter-shaped marks. The method 1. Prompt for the object with a blank surface: "a wooden shop sign, blank, no text", and put text, letters, words, watermark in the negative. 2. Render, pick the image. 3. Open it in an editor (GIMP, Krita, Photopea, Inkscape for vector). 4. Set the text in a real font. Match perspective with a perspective or distort tool. 5. Sell it: lower opacity slightly, add a matching shadow, and use a blend mode (Multiply on paper or wood, Overlay on painted walls) so the grain shows through. 6. Optional: run a very light img2img pass (denoise 0.15-0.25) over the whole image to unify lighting. Check the letters survived; if not, mask them out of the pass. When the text should be part of the art For hand-lettered or carved looks, set the word in a font, render it as a white-on-black image, and use it as a canny or depth ControlNet input. The model then paints the letters you supplied. Ancient scripts Never trust a model with hieroglyphs or cuneiform. See the Type, Scripts & Symbols board for real Unicode fonts and the Studio Scripts tool.