Kling

A Kling prompt starts by setting the subject, movement, and scene, and splits the description into several short sentences. In Kling Omni, you call connected reference images with numbered tags to tell the model what role each image plays.

Prompt structure

Write the subject, movement, and scene first, then add camera, lighting, and atmosphere to make the result more polished.

Kling 2.6, Kling 3.0, and Kling Omni use the same prompt structure.

Subject(Subject Description) + Subject Movement + Scene(Scene Description) + (Camera Language + Lighting + Atmosphere)

Always include the subject, movement, and scene.
Add the parenthesized camera language, lighting, and atmosphere when you want a more polished result.
It is best to write with simple words and sentence structures rather than elaborate expressions.

Rather than packing the whole description into one long sentence, split it into several short sentences. When everything goes into one sentence, the details can blur.

Shot in medium range, with a blurred background and atmospheric lighting, a giant panda, adorned with black-rimmed glasses, is reading a book in a café. The book lies on a table, accompanied by a steaming cup of coffee, next to the cafe windows, movie-level color palette.

It is best to put only one movement in each clip. When one clip holds several events, each event may not come through fully.

Prompts can be up to 2,500 characters. There is no negative prompt field for listing elements to exclude.

You can check the supported inputs, length, ratio, and resolution in Video models.

Adding dialogue

In Kling 3.0 and Kling Omni, pairing each line with the character who speaks it assigns the dialogue to each character.

Describe the character who speaks, then write that character's line right after. When characters and lines are paired, each line is assigned to its character automatically.

The woman in the navy coat says: "You're late again."
The man across from her replies, without looking up: "Traffic."

Dialogue can be written in Chinese, English, Japanese, Korean, or Spanish. Dialogue written in any other language is translated into English when the video is generated.

Dialogue is turned into speech when you generate with sound on. When a video input is connected to Kling Omni, no new audio is generated, so dialogue written in the prompt is not turned into speech either.

Calling reference images

In Kling Omni, you call connected reference images in the prompt with tags numbered in the order they were connected.

Reference images are numbered in the order you connect them to the video node. In the prompt, write each number as a tag wrapped in triple angle brackets.

Reference image	Prompt tag
First image connected	`<<<image_1>>>`
Second image connected	`<<<image_2>>>`
Nth image connected	`<<<image_N>>>`

<<<image_1>>> walks into a coffee shop and sits down at a table near the window. The interior matches the style of <<<image_2>>>.

You can connect up to 7 reference images, so tags go up to <<<image_7>>>. With a video input also connected, reference images are limited to 4, and tags go up to <<<image_4>>>.

Dividing roles between images and the prompt

The prompt carries the movement and flow that images cannot convey, rather than repeating what the reference images already show.

The most common mistake is describing again in the prompt what the reference images already show. Repeating the outfit, face, lighting, or background in words makes it unclear whether to follow the image or the text, and the result can end up further off.

In the prompt, write what the images cannot convey.

What moves
How the camera moves
What happens, and when

When you connect several images, first state what each image is for, then describe the scene.

<<<image_1>>> is the character. <<<image_2>>> is the location.
<<<image_1>>> sits at the counter in <<<image_2>>>, turning toward the door as it opens.

What to keep in mind when writing

These are elements Kling models have trouble rendering accurately, and ways to write so the result comes closer to what you want.

Counts may not be reflected exactly. Even if you specify a number, as in "10 puppies," that many may not appear.
Complex physical motion is hard to render. A bouncing ball or the path of an object thrown high may not be rendered accurately.
It is best to specify only one camera movement per shot. Asking for a push in, a pull out, and a rotation at the same time can make the frame shake.
On-screen text is not perfect. To add a sign or a caption, it is best to keep the wording short and state its position clearly.
To create an East Asian atmosphere or characters, words such as Oriental mood, China, and Asia are effective.