Can AI video models read your language? Prompt and dialogue languages, model by model
Most AI video and image models read English best, and several say so outright: Veo and Gemini Omni are documented for English, Wan and Qwen-Image for English and Chinese. What characters can speak is a separate list per family. Here is both, for every family on LUVI.

There are two different language questions behind every AI video prompt, and they have different answers. The first is which language the model reads its instructions in: on most families that is English, several vendors say so outright, and only Seedance 2.5 officially reads a long list of other languages. The second is which languages a character can speak with matching lip movement, and that list differs on every family—from three languages on Vidu Q3 to eleven stable ones on MiniMax H3. The rule that follows works everywhere: write the instruction in English, keep the spoken line or the on-screen text in its own language inside quotes, and pick the model by the language of the line. This post maps both questions for the video and image families in the LUVI catalog, as of October 20, 2026.
Which language should the prompt itself be in?
English, on every family where the vendor or LUVI's prompting guide says anything, with one wide exception. What the vendors document:
- Gemini Omni 1.1 Flash: Google's Omni documentation says "English (EN) is fully supported, but other languages have not been evaluated, so they may work but results can vary." LUVI's guide accordingly says to write the prompt in English "even when the spoken line or on-screen string is not".
- Veo 3.1: Google Cloud's model page lists "Prompt languages" as English.
- Wan: Alibaba's API reference for Wan 3.0 says the prompt "Supports Chinese and English".
- Qwen-Image: Alibaba's reference says the prompt "Supports Chinese and English."
- Seedance 2.5: the exception. ByteDance's ModelArk documentation says all Seedance models support English prompts (and Chinese, per its Chinese edition), and that Seedance 2.5 "additionally supports Spanish, Indonesian, Portuguese, Japanese, Malay, Thai, Arabic, Vietnamese, and Korean." French, German and Turkish are not on that list.
- HappyHorse: Alibaba's reference says the prompt "Supports any language", within a length limit.
- Kling 3, PixVerse, Grok Imagine, Hailuo, Seedream, GPT Image, Nano Banana: we found no vendor statement on prompt languages, and every example in LUVI's guides for them is English.
None of this means a Turkish, French or German prompt fails. It means the vendor has tested English, and a prompt in another language is a bet on how well the model generalizes.
Which languages can characters speak?
Spoken dialogue is a separate list for each family that generates sound, and a line outside the list tends to come back with weak lip-sync or in the wrong language:
- MiniMax H3: MiniMax lists "Stable support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish", with others "supported to varying degrees". LUVI's guide treats Turkish as outside the stable set and says never to invent a language tag for it.
- FLUX 3 Video: Black Forest Labs names English, Chinese, Spanish, French, German, Japanese, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi and Punjabi "and more, with precise lip-syncing". LUVI's guide adds: label each spoken line's language and keep it short.
- Kling 3.0: Kuaishou's launch announcement says the model "can generate speech in English, Chinese, Japanese, Korean, Spanish", plus American, British and Indian accents.
- Vidu Q3: Vidu's product page: "Generate videos in English, Japanese, or Chinese." LUVI's guide says another language "may render poorly—keep such a line very short".
- HappyHorse: LUVI's guide lists lip-sync in English, Mandarin, Cantonese, Japanese, Korean, German and French, with the line in its native script, and notes that "the prompt's own language does NOT drive the speech; the tag and the quoted script do." Alibaba publishes no list we could verify.
- Seedance 2.5: a line that is neither Chinese nor English must be tagged with its language—ByteDance's own example is
says in Japanese {こんにちは}. LUVI's guide adds that English lines need "in English" too, "or the line tends to come back spoken in Chinese". - Hailuo generates no sound at all, so it has no dialogue language. Grok Imagine Video's documentation names no languages.
For Turkish this narrows the field: FLUX 3 Video is the one family whose vendor names Turkish. On MiniMax H3, LUVI's guide suggests writing the shot without a dialogue block, describing the performance physically, and laying the voice in afterwards.
What about text inside the image?
On-screen text is copied, not translated, so it goes in quotes exactly as it must read, in its own script—and special characters are where models slip:
- GPT Image 2.5: LUVI's guide says to "keep non-English characters exact (Turkish İ ı ş ğ ç ö ü, ₺)."
- Nano Banana 2.1: non-English copy is kept "character-exact and LARGE in the frame; small lettering drops diacritics at any resolution."
- Seedream: on-image copy goes in double quotes in the target script, with the actual glyphs, never romanized.
- Qwen-Image: the quoted string keeps its language and capitalization; Alibaba's Qwen-Image repository notes its text rendering is strongest "especially for Chinese".
How do you write a prompt for a line in another language?
Keep the instruction in English and let only the quoted words change language. On MiniMax H3 the language goes inside a tagged dialogue block: <d>[French] Je descends au prochain arrêt.</d>. On Seedance 2.5 it is the language plus the line in its dialogue brackets: says in Japanese {こんにちは}. On HappyHorse it is a language tag and the line in quotes: says calmly in German, "Wir sind gleich da."
The structure is the same on every family: direction, camera and sound in English; the words a character says or a sign displays in their own language and script, inside quotes.
How we checked
- Models: the video and image families in LUVI's catalog that have a prompting guide.
- What was read: every guide's language rules for prompts, dialogue and on-screen text, the rules of LUVI's feather rewrite, and each vendor's own documentation where it states a language (linked below).
- Date and runs: October 20, 2026.
- Results: No outputs were generated for this post, so it makes no claim about output quality.
- Drawbacks: PixVerse, Grok Imagine Video and Hailuo publish no language list we could verify, so this post makes no claim about their dialogue languages. HappyHorse's lip-sync list comes from LUVI's guide, not from Alibaba. Vendors add languages without notice.
In LUVI
You can type your prompt in Turkish, Spanish, French or German in the Workspace. The feather button next to the prompt can rewrite it into the selected model's dialect, and the rewrite comes back in English "because that is what the models respond to best", while names and quoted lines keep the language you wrote them in (what the feather button does). For the sound side, see which models generate sound with the video; for a clip you already have, our workflow for dubbing a video into another language.
New to LUVI? Create a LUVI account and try one short line in your language on a family that lists it.
Sources
- Generate and edit videos with Gemini Omni Flash — Google AI for Developers, API documentation. Accessed October 20, 2026.
- Veo 3.1 — Google Cloud, model documentation. Accessed October 20, 2026.
- Wan2.7 text-to-video API reference — Alibaba Cloud Model Studio, API reference. Accessed October 20, 2026.
- Qwen-Image text-to-image API — Alibaba Cloud Model Studio, API reference. Accessed October 20, 2026.
- Qwen-Image repository — Qwen team, GitHub. Accessed October 20, 2026.
- Create video generation task — BytePlus ModelArk, API reference. Accessed October 20, 2026.
- Seedance 2.0 prompt guide — BytePlus ModelArk, prompt guide. Accessed October 20, 2026.
- HappyHorse text-to-video API reference — Alibaba Cloud Model Studio, API reference. Accessed October 20, 2026.
- MiniMax-H3 — MiniMax, GitHub repository. Accessed October 20, 2026.
- FLUX 3 Video, Part 1: Generation — Black Forest Labs, blog. Accessed October 20, 2026.
- Kling AI launches 3.0 model — Kuaishou Technology, press release via PR Newswire. Accessed October 20, 2026.
- Vidu Q3 — Vidu, product page. Accessed October 20, 2026.
More from Comparisons
- GPT Image 2.5 Flare vs Sunburst: same price, same settings—so what actually differs?On LUVI, GPT Image 2.5 Flare and Sunburst share every setting and quote the same credits at every tier, from 30 an image. What differs is time and judgement: Flare finished faster in LUVI's log, Sunburst read layout intent better in our tests. Pick by the job.
- HI3D vs Hunyuan3D: which image-to-3D model to run on a product photoBoth turn one photo into a GLB mesh, and neither reads a prompt on the image path. Hunyuan3D is the quick draft, from 600 credits, and the only one with text-to-3D. HI3D takes up to four views of the object and runs from 850 to 10,370 credits, whatever you switch on.
- Nano Banana, 2, 2.1 or Pro: which of the eleven Nano Banana models should you pick?LUVI lists eleven Nano Banana models across four generations, from 76 to 400 credits. Nano Banana 2.1 is the everyday choice at 80; Pro is for five faces from photos or a style reference; Pro Ultra is the 4k and 8k specialist and, at 4k, costs less than regular Pro.