Kling 4.0 Reference to Video: Cast Once, Shoot Every Scene

Getting one great shot of a character is easy now. Getting the same character in the second shot is where the afternoon goes. The jacket changes color, the jawline softens, the product label grows a letter it never had. You end up regenerating until something matches, and calling that a workflow.

Kling 4.0 Reference to Video is Kling's answer to that problem, and it went live on each::labs on October 24. Instead of describing your cast in words and hoping, you hand the model the cast: up to 10 images, up to 5 reference videos, saved Kling Elements and voices, as many as 15 references in one request. This guide is about using that budget well. Not every field (our Kling 4.0 Video Edit guide already walks through the shared inputs), but the habit that makes a series of shots look like it came from one production.

What Kling 4.0 Reference to Video Is For

Kling calls the feature Omni Reference, and its own user guide describes it as the default entry point with the most complete set of capabilities. The pitch on Kling's launch blog is consistency: the same woman across a table-side shot, a bedroom shot and an overhead shot with her face, hair and clothes still recognizable, or a product shown from different angles and lighting while its structure, materials, colors and brand details hold.

That makes it a different tool from the rest of the family. Text to video invents everything. Image to video animates a still along a path you plan with keyframes. Video Edit changes footage you already approved. Reference to video sits in the middle: nothing is locked except the things you showed it, and everything you showed it is supposed to stay put. If you're making one clip, you probably don't need it. If you're making the fourth clip of a campaign, you do.

On each::labs the endpoint only requires a prompt. Everything else is optional steering: images, videos, elements, voices, plus duration (3 to 30 seconds, 10 by default), resolution (720p or 1080p) and aspect_ratio (21:9, 16:9, 9:16, 1:1 or adaptive, which is the default).

Build a Casting Sheet Before You Write a Prompt

Here's the mistake almost everyone makes on day one. They write the prompt first and then go looking for images to attach. Flip it. Decide who and what has to stay identical across the series, then build one reference for each of those things, and only then write.

A casting sheet for a small skincare campaign might be four images: the presenter's face and outfit on a plain background, the bottle on white, the bathroom set with the light you want, and a storyboard frame. Four references, four jobs. That's the right density. Kling's image limit is 10 and the overall cap is 15, but the budget isn't a target. Every extra reference is another thing the model has to reconcile, and two photos of the same presenter in different lighting will fight each other.

Make the references boring on purpose. A clean, evenly lit image of the jacket tells the model about the jacket. A moody street photo of the jacket tells it about the street. Each image goes in as a refer_image with an id, and the schema asks that ids are unique across images, videos, Elements and voices. Kling's own interface numbers assets in upload order (@Image1, @Video1). On each::labs you choose the names, so pick ones that read like a call sheet: presenter, bottle, bathroom.

Give Every Reference One Job

Kling's user guide is blunt about this: if you provide asset ids without saying what they're for, "the model will infer their intended use, making the results less predictable." So the first lines of a good Kling 4.0 reference prompt aren't the scene. They're the role list.

bathroom defines the location. Preserve its layout, tiles and the soft window light from the left. presenter defines the woman. Preserve her face, short dark hair and cream knit sweater. bottle defines the product. Keep its shape, amber glass and label exactly as shown.

That pattern comes straight from Kling's examples ("@Image1 defines the scene. Preserve its spatial layout, furnishings, and lighting."), and it works because it does two things at once. It assigns a job, and it scopes what carries over. The presenter image gives you her face, hair and sweater. It does not give you the grey studio wall behind her, because you didn't ask for it.

When two references overlap, settle the argument in the prompt. If the presenter's photo shows her holding a different product, say which reference wins for the bottle. Kling's guide recommends defining each asset's scope and explicitly excluding anything that shouldn't carry over. Write it plainly: "Use presenter for her appearance only."

Write the Shot as a Timeline

After the role list, Kling's recommended structure is overall settings (scene, texture, visual style, camera tone), then a sequence of events with time stamps, then audio. The time stamps matter more than they look. Kling says multi-shot is on by default and the model splits shots and sets pacing by itself. That's great when you don't care. When you do, give it the beats.

Warm, natural, editorial look, 35mm feel, soft contrast. 0 to 4 s: close-up of bottle on the bathroom shelf, slow push in. 4 to 9 s: medium shot, presenter picks up the bottle, turns it so the label faces camera, smiles. 9 to 12 s: overhead shot of her hands pressing one drop onto her palm. Audio: quiet room tone, a single soft tap of glass on tile, no background music, no subtitles.

Notice the ending. Kling's guide warns that negative instructions "can increase the likelihood of errors," and names only two exclusions worth writing: no subtitles and no background music. So describe what you want instead of listing what you fear. "Clean label, sharp lettering" beats "no warped text." The same guide asks for concrete visual language over abstract praise, so "premium look" or "vintage feel" should become the actual light, lens and texture you mean.

Kling 4.0 prompts can run long (the schema accepts up to 26,000 characters in English, and Kling quotes 8,000 tokens), but length isn't the goal. Structure is. A role list, one line of look, three time-stamped beats and an audio line is usually enough for a 12 second shot.

Borrow Motion and Camera From a Reference Video

Images carry identity. Videos carry time. A feature_video lets you show Kling 4.0 how something should move without inheriting what it looks like. Kling's examples are short and specific: "Reference the camera path and pacing of @Video1," or "Reference the fighting moves and attack timing in @Video1," followed by which images define the characters.

This is the most useful trick in the endpoint for anyone with a phone. Film a rough pass of the move you want: walk a handheld arc around a chair, or have a colleague mime the gesture. Upload it as a reference, give it an id like camera_path, and say exactly what to take from it. Kling also lists white-model and depth-map clips as supported video references, which means a grey 3D blockout can describe a camera move and blocking without contributing a single texture.

Keep the budget in mind. Reference videos, a base video and video Elements share one quota: up to 5 clips and 30 seconds combined, with each clip 3 to 30 seconds long. If your series needs one camera move and one performance, that's two clips, and you still have room. If you want to keep the motion of a specific clip exactly, that's no longer reference work. It's an edit, and Video Edit is the better door.

Lock Recurring Characters With Elements and Voices

A photo is a good reference for one shoot. A character you'll use for months deserves an Element. Kling Elements are saved identities built in Kling, and the each::labs endpoint takes them by element_id, up to 7 per request with at most 3 video Elements. Kling says a video Element is built from 3 to 8 seconds of footage, and it counts against the shared video quota.

Voices work the same way. You can attach up to 3 existing custom or preset voices by voice_id, each with its own id for the prompt. Kling's guide adds two dialogue rules worth following. Write the spoken line in quotation marks in the language you want it spoken in, and describe the speaker's age, tone, speed and emotion. According to Kling's launch blog, 4.0 outputs two-channel stereo audio with better lip-sync accuracy, so a talking shot doesn't need a separate pass.

host speaks in a warm, unhurried voice, woman in her thirties, slightly amused: "One drop. That's the whole routine." Use voice_a for host.

If a character appears in more than a handful of shots, promote them to an Element. If the brand has a spokesperson voice, save it once and stop describing it in prose.

A Kling 4 Reference Workflow for a Five-Shot Series

Start with the casting sheet: one clean image per recurring person, product and location, each with a readable id. Write the role list once and save it as a text snippet, because every shot in the series will open with the same lines. That repetition is the consistency system.

Next, draft every shot as a timeline under that shared header, and decide per shot whether it needs a feature_video for camera or performance. Run all five at 720p with short durations to check that faces, products and lettering hold. When a shot drifts, don't add adjectives. Tighten the role line for whatever slipped ("Keep the label exactly as shown in bottle") and rerun.

Once the series reads as one production, rerun the winners at 1080p and at the final durations. Processing is asynchronous, so poll or wait in your integration rather than blocking. If you're generating variants for several markets, put the steps in an each::labs flow so the role list and settings stay identical every time. And if one approved shot only needs a different product or wardrobe, send it through Kling 4.0 Video Edit instead of regenerating it.

Reference, Keyframes or Plain Text: Picking the Right Kling 4.0 Door

Kling's guide states that Omni Reference and keyframes can't be combined in a single generation on its own interface. Treat them as separate tools. If you have a first frame and a last frame and want the model to fill the path between, use Kling 4.0 Image to Video, which takes 1 to 10 frames typed as first, middle or last. If you only have an idea, start with Kling 4.0 Text to Video. If you have a cast that must survive across shots, reference to video is the one.

And if you're still deciding whether Kling 4.0 is the right model for the job at all, our Kling 4.0 vs Kling 3 breakdown covers what changed, and the wider image-to-video catalog is one model id away. For a second take on the same problem from a different model, our MiniMax H3 reference to video walkthrough is worth a read.

Frequently Asked Questions

How many references can Kling 4.0 reference to video take?

Fifteen in total across every type. Within that, the each::labs endpoint allows up to 10 images, 5 videos, 7 Elements (3 of them video Elements) and 3 voices, and Kling caps the video pool at 30 seconds combined. In practice, four to six well-chosen references beat fifteen loose ones.

How do I refer to a specific image in my Kling 4 prompt?

Give it an id in the request and use that name in the prompt, then say what it controls: "presenter defines the woman. Preserve her face and sweater." Kling's own app uses names like @Image1 in upload order. The principle is the same: a reference with no stated role is a guess.

Is Kling 4.0 reference to video the same as Video Edit?

They share inputs, not intent. Reference to video only needs a prompt and builds a new shot around your cast. Video Edit requires a base clip and keeps its motion while changing what's in it. Generate the series with references, then use Video Edit for variants of an approved take.

What resolution and length does it support on each::labs?

Shots from 3 to 30 seconds in one-second steps, at 720p or 1080p, in 21:9, 16:9, 9:16, 1:1 or adaptive framing. Kling lists 4K and 10-bit HDR for Kling 4.0 on its own platform, with HDR marked as coming soon.