Kling 4.0 Omni Reference 指南：實際限制、提示詞結構與各段影片的參考照片

Kling 4.0 Omni Reference takes several pictures of one person or product at once and builds a new scene that keeps all of them: the face, the hair, the jacket, the label on the bottle. It reached everyone in Kling's own app on October 15, 2026, and its API opened on fal.ai and Runway Dev on October 23. We run it in RefVid, so this post is what we learned wiring it up: the limits that actually apply, how Kling wants the prompt written, and real clips next to the photos that made them. The full, regularly updated version lives on our Kling 4.0 Omni Reference guide.

Kling 4 in RefVid, from two reference photos of Mia, an invented character: 3 seconds, 720p, with sound.

Why a reference model, not image-to-video

Image-to-video animates one picture as the first frame. That is fine for one clip, but ask for a second scene and the model reinvents the subject: the jawline drifts, the hair parting flips, the logo on the mug changes shape. A reference model reads a small set of images as a description of who or what must appear, then builds the shot around that description. The more angles you give it, the less it has to guess.

One of Mia's four reference photos. The set: front, three-quarter, profile and full body, even light, plain wall.

Kling 4.0 vs Kling 3.0 Omni: the limits

From Kling's 4.0 user guide (updated November 4, 2026), the Kling 3.0 Omni guide, fal.ai's reference-to-video schema and Runway Dev's model list:

Per file, Kling 4.0 allows images up to 30 MB with an aspect ratio between 2:5 and 5:2, videos of 3 to 30 seconds up to 300 MB, and voice profiles made from 5 to 30 seconds of audio. Music and sound effects cannot be uploaded: audio input is voice profiles only, and the model makes the rest from the prompt.

Two things that surprised us. First, keyframes: in Kling's app, Keyframes (up to 10 images that steer how a shot changes) is a separate entry point and cannot be combined with Omni Reference in one task, but fal.ai's reference-to-video endpoint accepts a start frame, an end frame and up to 8 keyframes next to the references, all counted within its 10 images. Second, Kling 4.0 Flash also has Omni Reference, but only for 3 to 20 seconds at 720p, without start and end frames or keyframes.

What it costs per second

• fal.ai: $0.157 per second at 720p, $0.235 at 1080p, so a 10-second clip is $1.57 or $2.35; with video references, $0.094 and $0.141.
• Runway Dev: from $0.157 per second, up to 30 seconds at 720p or 1080p.
• Kling's own developer platform: early access through a business inquiry, preview prices $0.14 and $0.21.
• RefVid: 58 credits per second at 720p and 86 at 1080p on Kling 4, sound included; 31 credits per second on Kling 3.0 Omni.

If you are still choosing a tool rather than a model, we keep a side-by-side of ten image-to-video tools, sorted by the price of the cheapest paid plan, with clip lengths and how many pictures each one accepts.

How Kling wants the prompt written

Kling numbers your references in upload order: @Image1, @Image2, @Video1. Its guide asks you to give each one a job, for example:

@Image1 defines the woman; keep her face, hair and clothes.
@Video1: follow its camera path.

With bare tags and no jobs, the model guesses what each reference is for. The guide's prompt has four blocks:

1. Asset roles: what each reference defines and what to keep.
2. Overall settings: scene, texture, style, camera tone.
3. Sequence of events: what happens from 0 to X seconds, X to Y, and up to the end.
4. Audio: dialogue in quotes, ambient sound, music, effects.

Concrete details beat abstract words: "side light on a matte surface", not "premium look". Use negatives only for the two things the model adds on its own, subtitles and background music. To change an existing clip instead of making a new one, start with "Edit @Video1" and end with "Keep everything else unchanged".

Two examples: the photos, the prompt, the clip

In RefVid you type your subject's saved name and we swap it for the right @Element tag before the prompt reaches Kling 4. Below, {1} marks where the subject goes. Neither prompt describes the product; the photos carry its shape, finish and colours.

ATLAS, an AI-made watch, from 2 photos. Prompt, word for word:

Extreme macro of {1}: the camera glides slowly across the dial as the seconds hand sweeps, light rolling over the polished case and the hour markers, then pulls back to show the whole watch resting on dark leather. Warm low-key light, shallow depth of field, luxurious detail.

Reference 1 of 2: the dial straight on and glare-free, so the markers, hands and logo are legible.

The result: 5 seconds, square, with sound, on Kling 4. This is our watch macro template, which opens the studio with this prompt already filled in.

MARA, an AI-made handbag, from 2 photos.

A woman in a cream trench coat walks down a sunlit city street carrying {1} on her arm; the camera tracks alongside at hip height so {1} stays centred and sharp, swinging gently with each step, her face out of frame. Golden-hour light, shallow depth of field, fashion-ad style.

Reference 1 of 2: the front, plain backdrop, hardware and logo plate sharp.

5 seconds, vertical 9:16, with sound, on Kling 4: the handbag street template.

The pattern in both: where we are, what the subject does, then light, lens and mood in a few words. "Face out of frame" in the second one is deliberate: the bag is the subject, and an unreferenced face is one more thing for the model to invent.

Preparing references that work

These rules hold for Kling 3.0 Omni and Kling 4.0 alike, so a set prepared for one works on the other.

• Three to five images. A person: a clear front view, a three-quarter view and a profile. A product: the hero angle, the back or side, and a close-up of the logo or texture.
• Even light, simple background. A plain wall helps the model separate the subject from the scene; busy backgrounds leak into the output.
• The same subject in every image. Same outfit, same hairstyle, same colourway. Two outfits means two characters.
• No filters, watermarks or text overlays. The model treats them as part of the subject.

Where it still slips

Small text on products, such as ingredient lists and fine print, is often redrawn. Hands handling small objects can still look wrong. Angles that none of your references show, like the back of a head or the underside of a shoe, are invented rather than copied. The fix is almost always better references, not a longer prompt.

What it is good at: the same face, scene after scene

Where a reference model earns its cost is anything viewers watch more than once: a presenter who opens every tutorial, a product shown in five rooms, a cast that has to survive the cut between scenes.

The same four photos of Mia, three scene prompts, the same face each time (Kling 3.0 Omni in RefVid, before we switched the default to Kling 4).

That property is what a short series needs. We wrote up how to keep the same actors in every episode of an AI drama series: casting from saved photos, formats that suit one-to-two-minute vertical episodes, and writing scenes with the dialogue in quotes so the model can shoot them.

How we use it in RefVid

You save a character or product once from 1 to 10 photos. It goes to Kling 4 as one element: the first photo as the front view, up to three more as other angles, the rest as extra reference images. Then every clip starts from that saved subject; you write the scene, pick a length and a shape (16:9, 9:16 or 1:1), and generate. Kling 4 is the default and Kling 3.0 Omni stays as the cheaper choice; subjects saved before the switch carried over unchanged. New accounts get one free 5-second video, on Kling 3.0 Omni (Vidu Q2 in some regions), no card needed.

The full guide has the source for every number above, a third prompt example and an FAQ, and we update it when Kling or fal change a limit.

— Louis Chi, RefVid