Why image models invent fake hieroglyphs and runes, and how to check a render

Ask any image model for "a wall of hieroglyphs" and you get convincing nonsense: signs that look Egyptian but are not in any sign list, mirrored mid-line, repeated in patterns no scribe would use. The same happens with runes, cuneiform and Chinese characters. Why The model learned what these scripts look like as texture. Nothing in training tied a mark to a sign value, so it reproduces the statistics of the surface: density, line weight, a few frequent shapes such as the owl or the eye. Tokenisers also break rare words into pieces, so prompting the name of a sign does not help. How to check a render 1. Pick five signs at random in the image. 2. Look each one up in a sign list or the Unicode code chart for the script: https://www.unicode.org/charts/ 3. If you cannot find them, the inscription is decorative noise. 4. Check direction: Egyptian signs face the start of the line; mixed directions in one column are a giveaway. What to do instead - Leave the surface blank in the render ("smooth stone, no inscriptions") and set real signs from a Noto font, then composite. See the Unicode blocks thread on this board. - For carved or painted results, use the typeset signs as a canny or depth ControlNet map, so the model paints your signs. - Hathor Studio's Scripts tool is at https://hathor.soapbox.community/scripts When decorative is fine For a fantasy world, invented glyphs are a design choice. Say so in the post. The problem is only fake script presented as a real one.