A practical note for anyone making posters, covers or banners with text in them. Short text rendering has improved a lot in recent open models. I’ve mostly been using MOSS-Image (open weights, Apache-2.0), which handles both English and Chinese, and these habits make the biggest difference:

• Quote every string exactly. Title text at the top: "NIGHT SIGNAL" works far better than “a poster titled night signal”.
• One short string per block. A title, a subtitle and a date render cleanly; a full sentence usually picks up a typo somewhere.
• Give each string a position. Top, bottom, left third, centred. If the model has to guess where the text goes, it also tends to guess what it says.
• Name a type style, not a font. “bold condensed sans-serif” or “hand-lettered brush script” works; font names mostly get ignored.
• Bilingual layouts work if you separate the languages. For example, the English title at the top and the Chinese subtitle under it as two quoted strings, rather than mixing both in one line.
• Use enough steps. Small type is the first thing to break at low step counts. 50+ steps and guidance in the middle of the 3-6 range is a good baseline.
• Fix the seed while iterating on wording. Otherwise you can’t tell whether a fix came from the prompt or from luck.

Disclosure: I’m on the team behind Moss Image, an independent browser front-end for this model (not affiliated with the model’s authors). It has free credits for new accounts if you’d rather test this without a local GPU: Learn More

If you have prompts where text still comes out garbled, post them. I’d like to know which cases still fail.