Guide AI & ML 6 min read

Kling video gets a negative prompt field on Tencent Cloud's API pages but not on Kling's generation pages

A negative prompt is a separate field listing what an image or video model should avoid, and Chinese video APIs document it endpoint by endpoint. As read on 16 November 2026, Tencent Cloud's pages for the Kling models it resells list one, Kling's own video-generation pages do not, Wan's pages list one and Tencent's own Hunyuan video pages do not.

Priya Nair
Data, AI Governance & Policy Analyst
Published 17 Nov 2026, 7:39 AM (SGT)
Share:
A camera operator's hand on a professional cinema camera rig during an indoor shoot. A camera operator's hand on a professional cinema camera rig during an indoor shoot. Photo by Luke Miller on Pexels
Advertisement

A negative prompt is a second text field that tells an image or video model what to steer away from, such as a watermark, blur or an extra finger; in Chinese it is the 反向提示词 (fǎnxiàng tíshìcí) or 负面提示词 (fùmiàn tíshìcí). Whether a Chinese provider's API has one depends on the endpoint, and for Kling on who sells it. As read on 16 November 2026, Tencent Cloud's pages for the Kling video models it resells document a NegativePrompt of up to 2,500 characters. None of the 20 video-generation reference pages on Kling's own API site lists a negative field. Kling's multi-element video editing endpoint does, also at 2,500. Alibaba's Wan documents a 500-character negative_prompt for text-to-video and image-to-video, and the pages for Tencent's own Hunyuan video endpoints list none.

Kling video checks in a negative prompt separately on Tencent Cloud's pages

We read each provider's own API reference on 16 November 2026 and searched every page for a negative-prompt parameter. Nothing was called, so what follows is what each page documents; it says nothing about what an endpoint does with the field. The video pages, counted per endpoint group:

Provider and productVideo pages readWith a negative fieldField and limit
Alibaba 万相 (Wan)text-to-video, image-to-video2 of 2negative_prompt, 500 characters
Tencent 混元生视频, Kling resaletext-to-video, image-to-video, video extension, Omni-Video, motion control3 of 5NegativePrompt, 2,500 characters
Tencent 混元生视频, Tencent's own models2 endpoints0 of 2none
Kuaishou 可灵 (Kling), video generation20 pages, 2.5 Turbo to 4.00 of 20none
Kling, video editingmulti-element video editing1 of 1negative_prompt, 2,500 characters

Tencent's Kling text-to-video page, updated 22 April 2026, gives NegativePrompt as 「负向文本提示词。不能超过2500个字符」 (negative text prompt, no more than 2,500 characters), with the example value 模糊 (blurry). Its model list runs from Kling-V1 to kling-v3, and the image-to-video and video extension pages carry the same field. Kling's own generation pages say instead, in wording that varies slightly from page to page, that the prompt itself can carry "positive and negative descriptions". Its multi-element video editing page, which carries no date, has negative_prompt that "Must not exceed 2,500 characters". The two sellers describe the same model family differently, and reading the pages cannot tell which description a given request meets.

Wan stows the field under input or parameters depending on the model

Alibaba's text-to-video reference (last modified 27 October 2026) puts negative_prompt under input for wan2.7-t2v. On the text-to-image V2 page, wan2.6 moves it under parameters, while wan2.5 and earlier keep it under input. Tencent's image API spells it NegativePrompt on two of the four Hunyuan image endpoints we read. Kling's image generation (kling-v3, kling-v2-1) has negative_prompt, but its page says that in image-to-image, "negative prompts are not supported".

Our Chinese negative prompt builder turns ticked flaws into one list per destination. Before a revision dated 16 November 2026, its single Chinese target was a generic field metered against Wan's 500-character cap, and its Midjourney line dropped every multi-word term, which removed 4 of its 8 default English terms. It also stripped the leading 无 from a word such as 无线耳机 (wireless earphones), leaving 线耳机. The revised builder offers Wan, both Hunyuan image endpoints, Kling and Tencent's Kling resale as separate targets, each metered against its documented limit. With its six default boxes ticked and the Tencent Kling resale chosen, it writes 模糊,低分辨率,噪点,多手指,畸形手,水印, 21 characters of 2,500.

Advertisement

Excess past 500 characters is cut off at Wan, and Tencent leaves its unit unstated

Wan's English edition of the text-to-image V2 reference says the field "Supports Chinese and English" and that "Excess characters are truncated automatically" past 500.

Tencent counts differently. SubmitHunyuanImageJob allows 「最多可传1024个 utf-8 字符」 and recommends Chinese; TextToImageLite allows 256 of the same unit; Tencent's own video endpoints cap the main prompt at 200. The phrase utf-8 字符 could mean characters or UTF-8 bytes, and the pages do not say which. A Han character is three bytes, so on our arithmetic a 100-character Chinese list is 300 bytes: inside 256 if Tencent counts characters, over it if it counts bytes. The revised builder shows both counts and warns on either.

ByteDance's Volcengine Ark pages give no cap at all for the prompt, only a recommendation (建议): under 300 Chinese characters for Seedream images and under 500 for Seedance video.

Providers with no hold for exclusions point writers back to the prompt itself

MiniMax's 14 generation references, on both its international and Chinese sites, list no negative field. Neither do Ark's image and video pages, Zhipu's three CogView and CogVideoX references, or the six 即梦AI (Jimeng) interface pages that rendered; four more did not render enough to read.

Where a vendor addresses the gap, the advice is to write it into the main prompt. Kling's Kling 4.0 manual, in its section on negative control, says to "Prefer positive descriptions of what you want to see", and to write "no subtitles" in the prompt itself if the model adds them. It also warns against a long register of unrelated exclusions. Tencent, even where it offers the field, says on its TextToImageLite page that a negative lowers the chance of the content appearing 「但不能完全杜绝」 (but cannot rule it out entirely).

Midjourney's moderation inspects each word after --no on its own

Midjourney's No parameter article, updated 27 July 2026, says its "moderation system" reads every word after --no independently, so --no modern clothing becomes no "modern" and no "clothing", which can trigger a warning. The same article equates --no with weighting part of a multi-prompt at −0.5. It is a statement about moderation, and it says nothing about the image model reading phrases word by word.

The revised builder keeps multi-word terms in its --no line and lists low resolution, extra fingers, mutated hands and bad hands under a moderation note, suggesting single words or a positive description if a warning appears.

Seedream and Seedance pages clear Thai and Malay prompts with no exclusion field

For ASEAN languages the question moves from the negative field to the prompt. Ark's image generation page (updated 13 November 2026) names Thai, Malay, Indonesian, Vietnamese and Filipino among the extra prompt languages for Seedream 5.5, 5.0 pro and 5.0 flash. The video generation page (updated 15 November 2026) lists Malay, Thai, Vietnamese and Indonesian among Seedance 2.5's languages, without Filipino. Neither page has a field for exclusions, so on these models a Thai or Malay prompt has to carry its own.

Wan's text-to-image page lists a Singapore region beside Beijing and Virginia, each with its own API key, and its negative field is documented for Chinese and English. No page we read says a negative field accepts Thai, Malay or Vietnamese. For how these modalities fit together more broadly, see our guide to image, video and voice models.

Advertisement
Priya Nair
Data, AI Governance & Policy Analyst

Priya Nair covers AI governance, data protection, privacy, and digital trust topics for RECATOOLS.

View author profile → · Editorial policy

About this byline Priya Nair is a RECATOOLS editorial persona for AI governance, privacy, and digital trust coverage. Articles are produced and reviewed under RECATOOLS editorial supervision.

Corrections policy

Advertisement