Kling 4 Omni Reference: one character, every shot
Kling's Omni models let you hand the video model several reference images at once — a face from the front, a three-quarter view, a close-up of a jacket — and ask for a new scene that keeps all of it. Kling 4 is the next step of that line. This page explains what reference-based generation is good at, where it still slips, and how RefVid uses it so you upload your references once instead of re-attaching them to every prompt.
What "Omni Reference" means
Most text-to-video and image-to-video tools take one picture as a starting frame. That works for a single clip, but the moment you ask for a second scene the model reinvents the character: the jawline drifts, the hair parting flips, the logo on the mug changes shape. A reference-capable model works differently. Instead of animating one frame, it reads a small set of images as a description of who or what must appear, then builds a new shot around that description.
Kling's Omni line accepts several reference images and short reference clips in one request (the current Omni model takes up to ten images and five videos). The more angles you give it, the less it has to guess. Kling 4 Omni Reference continues that approach with the newer generation of the model.
Where Kling 4 stands today
As of October 2026, Kling 4.0 is in early access. We have not seen a confirmed date for open API access, and we will not pretend otherwise. RefVid is built so the model underneath can change without you redoing anything: today generations run on Kling 3.0 Omni through fal.ai, and when Kling 4.0 becomes available through the API we switch the default and your saved characters and products carry over unchanged.
RefVid is an independent platform. We are not affiliated with Kuaishou or the Kling team; we use the model through a commercial API like any other customer.
What reference-based video is good at
Series content. A brand mascot that appears in a weekly short, a product shown in five different rooms, a presenter who opens every tutorial — anything where viewers notice when the subject changes between clips.
Ads and UGC-style variations. You can keep the same person holding the same product while testing different hooks, settings and camera moves, which is exactly what ad testing needs.
Story and anime sequences. Characters stay recognisable from scene to scene, so a short narrative does not fall apart the moment the camera cuts.
Where it still slips
Reference models are much better than single-image animation, but they are not perfect. Small text on products (ingredient lists, fine print) is often re-drawn. Hands interacting with small objects can still look wrong. Extreme angles that none of your references show — the back of a head, the underside of a shoe — are invented rather than copied. The fix is almost always better references, not a longer prompt.
How to prepare references that work
Use three to ten images. Include a clear front view, a three-quarter view and a profile for a person; for a product, the hero angle, the back or side, and one detail close-up of the logo or texture.
Keep lighting even and backgrounds simple. A plain wall or a neutral backdrop helps the model separate the subject from the scene. Busy backgrounds leak into the output.
Keep the subject the same across images. Same outfit, same hairstyle, same product colourway. If you want two outfits, make two characters.
Avoid heavy filters, watermarks and text overlays. The model treats them as part of the subject.
Add a short reference clip if you have one. A few seconds of the person turning their head or the product rotating gives the model information no still image can.
How RefVid uses it
In RefVid you create a Character or a Product once and upload its references. After that, every generation starts from that asset: you write the scene and the action, pick a length and an aspect ratio, and RefVid packs the references the way the selected model expects. Different models have different limits — Kling Omni takes up to ten images, Veo 3.1's reference mode takes three — and that bookkeeping is ours, not yours.
Every video you make with an asset is kept in that asset's series view, so you can see at a glance whether the face or the product still matches and make another clip with one click.
Frequently asked questions
- Is Kling 4 Omni Reference available now?
- Kling 4.0 is in early access as of October 2026, without a confirmed date for open API access. RefVid runs on Kling 3.0 Omni today and switches to Kling 4.0 when it becomes available through the API; your saved assets carry over.
- How many reference images should I upload?
- Three to ten. Cover the main angles of the subject with even lighting and a simple background. More angles mean less guessing by the model.
- Is RefVid made by Kling?
- No. RefVid is an independent platform that uses Kling and other video models through their commercial APIs. It is not affiliated with Kuaishou or the Kling team.