The lettering law

First sentence: the encoders in the popular text-to-image models "lack character-level input features, making it much harder to predict a word's visual makeup as a series of glyphs." That is Liu and colleagues at Google Research (ACL 2023), and their character-aware variants beat the character-blind ones by more than thirty points on rare words. Second sentence: when Glyph-ByT5 (ECCV 2024) pushed design-benchmark text accuracy from under twenty percent to nearly ninety, it did so by fine-tuning a byte-level text encoder on glyph-aligned data. The two requirements they name are "character awareness and alignment with glyphs." Both are properties of the encoder.

So a model whose encoder is token-based cannot be prompted into reliable lettering, for the same reason you cannot be talked into seeing ultraviolet. The competence has to be trained in. Until a vendor tells you, in the documentation and not the launch post, that it has changed the encoder, every word you ask it to spell is a coin it will eventually drop. On four thousand printed bags, eventually is the operative word.

Read on 18 September 2026, from the documentation pages and nowhere else.

• OpenAI, image generation guide: the model "can still struggle with precise text placement and clarity", and with "placing elements precisely in structured or layout-sensitive compositions." Two admissions in one guide, one about type and one about layout.
• Ideogram, whose entire pitch is typography, on its text-and-typography page: accuracy is best in English and non-Latin scripts "often produce unpredictable results"; longer text raises "spelling errors, distortions, or incomplete words"; and, decisively: "consider using Ideogram to generate the visual concept and then add the text manually using graphic editing tools afterward." The typography vendor recommends setting the type yourself.
• Midjourney: the Parameter List, Version, Style Reference, Edit Model, Omni Reference and Terms pages make no claim about rendering type and list no text control. That silence is taught as an absence, not a quote, and it is the most honest position of the three.
• Google is the counter-claim: the Gemini image docs promise "legible, stylized text for infographics, menus, diagrams, and marketing assets." Read it carefully. Legible is not spell-correct, it is not your brand's typeface, it is not an exact string, and it is not editable afterwards. If the claim is true it moves the failure rate; it does not move the six properties in Part 2.

Two sentences at the end of every brief, before the ratio. They are the difference between the fourth and fifth rungs on the bench below.

No text, no lettering, no logo anywhere in the image.
Keep the upper third as quiet negative space; the headline is set later in HTML.

The first clause is a negative constraint, and negative constraints persuade rather than bind, so it will occasionally be ignored; when it is, the asset is rejected on sight and regenerated, which costs a minute. The second clause is a composition instruction, which the reference kit from session 2 can reinforce with a composition reference that has an empty third. Together they mean the model is never asked to spell, so it never spells wrong, and the only spelling on the asset is the string a person typed.

... with the title "Autumn Roast" in elegant lettering across the top.

Eleven characters, and the bench grades the whole brief red for them. Not because the picture would be worse, but because the encoder cannot be trusted with them and the fix does not live in the prompt.

Two-layer assets are how print and web illustration have always worked. A magazine cover is a photograph and a type layer. A poster is an image and a typeset headline. Nobody asks the photographer to also paint the masthead. The model is the photographer here, and asking it to letter is asking the wrong craftsperson. The composite is not a patch for a weak model; it is the ordinary structure of a finished piece, and it is the structure that keeps every downstream job (translation, legal edits, accessibility, search) possible.

There is a second reason, which session 5 carries in one card: a picture where a person chose the composition, refused the lettering, set the type and arranged the layers is a picture with a documented human contribution in it. A picture that is one prompt and one PNG is not. The two-layer asset is the version that can be shown to have been made by someone.

• The first prompt: seven style words, zero anchors. "Beautiful", "trending on artstation", "masterpiece", "highly detailed", "8k", "stunning" and "illustration". Every word is about how good the picture should be and not one names an object, a light or a colour. There is no picture here for a model to make; there is a request to be impressed.
• Concept before art: fifteen anchors, two style words. A ceramic mug, a worn oak counter, three roasted beans and a green one, morning light from the left, amber and burnt orange and cream, flat vector with paper grain, 3:2. Fifteen things a painter could be wrong about. This is a picture.
• With the title spelled in: one added line, verdict red. Eleven characters inside quotation marks, and the brief that was a picture is now also a spelling test the encoder will eventually fail.
• Refused and reserved: the two clauses. Same fifteen anchors, same two style words, plus "no text, no lettering, no logo" and an empty upper third. The verdict goes to good. Nothing about the picture changed; what changed is that the model is never asked to do the one thing it cannot.
• More style words: the trap. Fourteen adjectives from the prompt guides bolted onto the concept brief. Sixteen style words across seven families, and four of them contradict: flat against photorealistic, flat against watercolour, watercolour against photorealistic, flat against highly detailed. Every anchor is still there. The picture did not get richer. The instruction became incoherent, and an incoherent instruction to a persuading layer is a coin toss per adjective.
• Wired to the kit: four reference handles. The refused brief plus the style reference set, its weight, a named asset to match, and a recorded seed with its ninety-nine percent written next to it. The verdict adds "wired to the kit". This is the brief that can be run again next month.

Paste a prompt you have actually used. Read the eight counts in order: anchors first, because a brief with fewer than five is not describing a picture yet; then style words and conflicts, because most real prompts have three to eight style words and at least one pair at war; then lettering, because most real prompts ask for some.

Then fix it in the order the counts suggest. Add anchors until you pass five: a material, a light, a placement, a colour. Cut style words until one family is left. Move every quoted string out of the brief and into a note that says "set in HTML". Add the two clauses. Watch the verdict move, and notice that at no point did you add a quality word.

If you finish early, take the concept brief and try to break the anchor count without touching the style words: remove "oak", remove "from the left", remove "3:2". The verdict degrades one step at a time, which is what a count is for.

A course that ships no generated image cannot grade one, and would not want to: a picture is judged by a person with the six sentences from session 2 in hand, and a browser scoring aesthetics would be inventing a number. What a browser can do honestly is read the brief the way the encoder reads it, as text, and count the things in it that are known to go wrong. That is a smaller claim and a true one, and it targets the layer where most people spend all their effort and none of their attention.