Marketing Oops tested five AiPASS image models and left both Seedream models unranked on Thai text

Marketing Oops ran the same prompt through all five image-generation models on Thailand's state AiPASS platform and reported that both Seedream models produced unreadable on-image Thai. In the test, published on 6 September 2026, GPT-Image 2.0 placed first, Nano Banana Pro second and Nano Banana third, while Seedream 4.0 and Seedream 5.0 Lite were left unranked.

This is one publication's hands-on impression, not a benchmark. One run per prompt, no seed control, no repeat trials, and the article itself notes that results can change with model version, settings and prompt style. What survives those limits is the method, and the method is the part worth copying.

Model choice for Thai creative has to be tested on rendered Thai text

The finding Marketing Oops reported on 6 September 2026 is that two of the five models on AiPASS failed on Thai characters specifically, not on image quality in general. Seedream 4.0 and Seedream 5.0 Lite went unranked because ad-image output did not follow the prompt as the team expected, on-image text was unreadable, and characters did not match Thai script.

That is a failure mode an English-language model comparison cannot detect. A global leaderboard, a demo reel or a vendor sample sheet built on English prompts will rank a model on composition, lighting and prompt adherence, and will say nothing about whether it can place a Thai word on a poster without mangling it. A Thai team picking a model from an English benchmark is picking on the wrong axis.

The rule that follows is short. Run your own Thai prompts through whatever models your tool gives you, and judge the Thai characters before you judge the aesthetics. An image that looks good and spells the product name wrong is not a draft, it is rework.

What TH-AI Passport is, and who can register

TH-AI Passport is a programme of Thailand's Ministry of Digital Economy and Society that gives Thai nationals free access to a centralised AI platform called AiPASS. Public registration opened on 19 August 2026 and the platform began operating on 31 August 2026.

Advance registrations passed 1.3 million in under two weeks, against up to five million available slots. AiPASS aggregates AI tools from more than 14 providers, covering more than 30 models, of which the five image models tested are one slice.

Eligibility is narrow in one respect and generous in another. A registrant must be a Thai national aged 15 or over on the date of registration, and must register personally, with no proxy registration allowed. Registration requires a Google, Microsoft or Apple account plus Thang Rath or ThaID for identity verification. Registering before 30 November 2026 upgrades the user to the Creator tier.

What Marketing Oops tested, and how they ran it

Marketing Oops used the AiPASS image-generation mode, which offers five models: Seedream 4.0, Seedream 5.0 Lite, Nano Banana, Nano Banana Pro and GPT-Image 2.0. They ran the same prompt through every model across five tasks drawn from everyday marketing work, which were a product ad image, a cafe food menu, an infographic, a product catalogue and a six-shot storyboard. The full write-up is on Marketing Oops.

Those five tasks are a fair spread of what an in-house team actually produces in a week, and they stress different things. An ad image tests composition and prompt adherence. A menu and an infographic test dense small text, which is where Thai rendering breaks first. A catalogue tests consistency across many items. A storyboard tests whether a character or a scene holds across six frames.

The ranking Marketing Oops reported on 6 September 2026

The table below is the outcome as Marketing Oops reported it, with the reason each model landed where it did. It is their test result, not an established property of the models.

Why Thai script is harder to render than English

Thai writing stacks marks vertically and runs words together, which gives an image model more ways to go wrong than a Latin alphabet does. Vowels can sit above, below, before or after a consonant, tone marks stack above vowels, and there are no spaces between words inside a sentence.

Analysis, not a claim from the source: a model generating text inside an image is drawing shapes it has learned, not typesetting with a font. Latin letters sit on one baseline in a fixed order, so a plausible-looking shape is usually a correct letter. Thai asks the model to place a mark at the right height above or below a specific consonant, and to do it without the word boundaries that would otherwise tell it where a cluster ends. A model can be excellent at English lettering and still produce Thai that reads as decorative squiggles, which is the outcome Marketing Oops described for both Seedream models.

It also explains why the menu and infographic tasks were the ones that exposed the weakness. Both put many short Thai strings at small sizes, where a misplaced tone mark is the difference between a real word and a shape.

A five-prompt Thai test set you can run in an afternoon

Nothing in the Marketing Oops test needs a research budget to repeat, and repeating it on your own brand is more useful than trusting anyone's ranking, theirs included. Use a fixed set of prompts, run them identically across every model your tool offers, and score Thai character accuracy first.

• A product ad image carrying your actual product name in Thai, at the size it would run.
• A price list or menu with at least eight Thai lines, which is the small dense text case.
• An infographic with Thai headings and Thai labels on the chart, not only a Thai title.
• A catalogue layout with the same product in several variants, to test consistency across items.
• A six-frame storyboard with one recurring character, to test whether the subject holds across frames.

Score each output in two passes. First pass: can a Thai reader read every character on the image, including tone marks and the product name. Anything that fails here is out, whatever it looks like. Second pass, only then: composition, brand fit, and how much prompt work it took to get there. Keep the prompts in a document so the test is repeatable when a model version changes, because model versions change and a ranking from September will not hold indefinitely.

What the Marketing Oops test did not establish

The write-up is a hands-on impression and does not claim otherwise, so the gaps are worth stating rather than papering over.

• One run per prompt, with no repeat trials and no seed control, so run-to-run variation is unmeasured.
• No sample size in any statistical sense, which means a single unlucky generation could account for a placing.
• No cost or quota detail for using the models through AiPASS.
• No comparison against paid direct access to the same models, so it does not say whether the platform or the model is the limiting factor.
• The article notes that results can change with model version, settings and prompt style, which is the honest caveat on any single-run test.

The licensing question the article does not answer

Marketing Oops did not address whether creative produced through a state platform is licensed for commercial brand use, and that is the question a brand team has to answer before any of this output ships. Nothing in the test speaks to it either way.

Analysis: the sequence that keeps a team out of trouble is to settle usage rights before the model comparison rather than after, because a model that wins your Thai text test is useless if the output cannot appear in a paid ad. Until that is confirmed in writing, the low-risk use of AiPASS output is internal work such as drafts, mood boards and layout tests, and anything customer-facing goes through a channel whose licence terms your team has already read.

What this means for marketers in Thailand

This is a Thai story with Thai facts, so little inference is needed. With advance registrations past 1.3 million and the platform operating since 31 August 2026, a large number of Thai marketing teams now have free access to the same five image models, which means the tools stop being a differentiator almost immediately. What separates the output is prompt discipline, the brand system behind it, and the willingness to reject a good-looking image with broken Thai on it.

There is a second split worth keeping clear. Being found and cited by AI answer engines is one job, and it belongs to AI search visibility work. Producing creative with AI models is a different job with different failure modes, and this story is entirely the second. Treating them as one thing is how teams end up measuring one and paying for the other.

For teams turning out a steady flow of visuals, the Thai character check belongs inside the content marketing workflow that briefs them and inside the web design process where they land, rather than as a separate review step that someone has to remember. Put it in the template, not in a person's head.

FAQ about the AiPASS image model test

Which image model on AiPASS handled Thai text best?

Marketing Oops placed GPT-Image 2.0 first in the test they published on 6 September 2026, calling it the most consistent across all five tasks even from thin prompts. That is their hands-on result from one run per prompt, not a benchmark.

Are the Seedream models bad at Thai?

Marketing Oops left Seedream 4.0 and Seedream 5.0 Lite unranked in their test because on-image text was unreadable and characters did not match Thai script. The test was a single run per prompt, so it records what happened in their session rather than establishing a fixed property of the models, and the article notes results can change with model version, settings and prompt style.

Who can register for TH-AI Passport?

A Thai national aged 15 or over on the date of registration, registering personally with no proxy allowed. Registration needs a Google, Microsoft or Apple account plus Thang Rath or ThaID for identity verification, and registering before 30 November 2026 upgrades the account to the Creator tier.

Can I use AiPASS images in a paid brand campaign?

The source did not state it. Marketing Oops did not address whether creative produced through the state platform is licensed for commercial brand use, so a brand team has to confirm usage rights independently before shipping anything.

Does the test say anything about cost?

No. The article gave no cost or quota detail for using the models through AiPASS, and made no comparison against paid direct access to the same models.

Where to go from here

If your team registered for TH-AI Passport in the first fortnight and is now deciding which model to standardise on, run the five Thai prompts above before you decide. If you would rather have the test designed around your own type, product names and channels, with the licensing question settled before anything ships, that is work we can set up with you. Tell us which models your tool gives you and we will start there.