Can a better prompt fix Arabic and Hindi text? Three tricks, three models, 12 GB card

Failed2026-10-27 · FLUX.3 klein, Qwen-Image 2.1, MOSS-Image SFT · RTX 5070 12GB

Earlier today Arabic text failed on all three models and Hindi failed on two. The obvious follow-up: does the way the prompt is written change that?

The test

We re-ran the failing strings with three changes, two seeds each, same settings as before, one rented RTX 5070 12GB per model (24 pictures per model): 1. Script hint: add "written in Arabic script" / "written in Devanagari script, spelled exactly as given". 2. One word only: ask for a single large word ("مهرجان", "विशेष"…) instead of two lines. 3. Plain card: one line of large bold black text on a white card, nothing else.

What we saw (judged by eye from the pictures, not counted)

• None of the three tricks fixed Arabic. Even on a plain white card, every model still changed letters in the first word of «طبق اليوم»: «طيق» (Qwen, FLUX), «طتى» (MOSS), and one-word menus came out as «طرق» or «نطيق». Titles on the card came out as «مجاجان», «مبهجان», «ميرحان» or «مرجل» instead of «مهرجان».
• Hindi stayed model-dependent. FLUX.3 klein again wrote «उत्सव» and «आज का विशेष» cleanly in most pictures; its «ग्रीष्म» conjunct still looks slightly off on some. Qwen kept writing «विशेथ» and MOSS «विशेए» even on the plain card.
• The only visible gain: Qwen wrote the menu line «हरी चाय» correctly in both seeds once we added the script hint (it had written «हरी याय» before).
• Plain cards did make the text easy to read, which makes the errors easy to see; they did not make the spelling correct.

How much to trust this

• We do not read Arabic or Hindi natively and compared letters with the prompt; no numeric score this time because several pictures were borderline. Corrections are welcome via the report page.
• Two seeds per prompt, one machine per model.
• It suggests these models need more than prompt wording for Arabic and some Hindi conjuncts; we have not tested fine-tuning or other tricks.