Kling V4.0 Reference-to-Video: images, videos and voices in one clip
Kling V4.0 Reference-to-Video mixes up to ten images, five videos and three audio clips into one new video. How the @ tags work, three recipes and a cheap test.
Most video models take one picture and imagine the rest. Kling V4.0 Reference-to-Video takes a whole folder: up to ten reference images, up to five reference videos and up to three audio clips, and builds a new clip of 3 to 30 seconds that uses them. The character from one picture, the jacket from another, the camera move from a phone video, the voice from a recording. It is the closest thing in the catalog to briefing a small film crew, and like any brief, it works only when it is clear who is who.
The one rule: name every reference in the prompt
Each upload gets a tag by its position: the first image is @Image1, the second @Image2, and so on; videos are @Video1 to @Video5, audio clips @Audio1 to @Audio3. The prompt is required, and it is where you say what each reference is for. A reference you upload but never mention is a guess for the model. Keep the order of uploads in mind, because swapping two pictures after writing the prompt silently swaps their roles.
Recipe 1: the same character in a new scene
Two images: a front view of your character and a detail of the outfit. Nothing else. This is the simplest way to see whether the model holds the face and the clothes before you add motion references.
@Image1 is the woman, @Image2 is her red raincoat. She walks out of a night market into light rain, pulls the hood of the raincoat up and looks back over her shoulder. Neon reflections on the wet street, slow tracking shot from the side.Recipe 2: borrow a camera move from a real video
Reference videos sit under the advanced settings. They guide motion, camera and pacing, so a ten-second phone clip of you walking around a table can become an orbit around a product that never existed. Together the videos are used for up to 30 seconds.
@Image1 is the ceramic teapot. Follow the camera path of @Video1: a slow orbit at table height, ending close on the spout. Morning light from a window, steam rising, a wooden table.There is also a switch called keep original sound. With it on, the clip keeps the sound of your reference video instead of generating new audio. It only works when you uploaded a video, and it is the right call when the reference already has the music or the ambience you want.
Recipe 3: a voice and a face
Reference audio clips, also under advanced, guide the voice and sound of the result. Each one should run 5 to 30 seconds. The clip comes out with synchronized stereo audio, so a character can speak in the tone of your recording inside a scene the model builds.
@Image1 is the old lighthouse keeper, @Image2 is the lighthouse. He stands at the top of the stairs with a lantern and speaks to the camera in the voice of @Audio1, slow and warm, wind and waves in the background.Settings that matter
- Resolution: 720p by default, 1080p on request. Draft at 720p.
- Duration: any whole number from 3 to 30 seconds; the default is 5. Longer clips give the model more time to drift away from the references, so earn the length with a good short version first.
- Aspect ratio: auto follows your reference images or video. Set 16:9, 9:16, 1:1 or 21:9 only when the destination needs it, and then give it references of a similar shape.
Where it falls short
Ten references is a ceiling, not a target. Pictures that disagree (two different hair colors for the same person, a daylight outfit and a night scene) give a result that averages them, and small details such as text and jewellery are where drift usually shows first. If you only need to animate one picture, the Kling V4.0 Image-to-Video model is simpler and gives you first and last frame control. For more on choosing references that agree, read reference images for a consistent style; for another take on the same job, see the Vidu Q4 reference-to-video guide. The full list is on the models page.
Questions
How do I tell the model which reference is which?
By tag in the prompt: @Image1, @Image2 and so on for images in upload order, @Video1 for videos and @Audio1 for audio clips.
How many references can I use?
Up to ten images, five videos and three audio clips. Videos and audio clips are under the advanced settings.
How long and how sharp can the clip be?
From 3 to 30 seconds, 5 by default, at 720p or 1080p, with synchronized stereo audio.
