Omni reference: what AI video models now take as input

"Reference" covers a family of inputs that tell a video model what to keep rather than what to invent. Where a text prompt describes a new thing, a reference hands the model an existing thing, a face, a product, a camera move, a voice, and asks it to carry that through the shot. Omni reference is the expanded form now shipping: one request that can hold images, video and audio references at the same time.

The idea moved to the front of the field on 15 October 2026, when Kuaishou's Kling 4.0 arrived with Omni Reference among its headline features. In the same week, MiniMax's H3, an open-weight distillation of it and an academic editing model all published reference-first modes. The count each model allows is now the number creators compare.

What Is a Reference, and How Is It Different From a Frame?

A first frame is where a shot starts. MiniMax's API documentation lists image-to-video as a separate mode built from a prompt plus a first-frame image and/or a last-frame image, and describes its job as controlling where a shot begins or ends. A model given a first frame starts on that image and animates forward. A last frame does the reverse, and a request that carries both runs between the two images.

A keyframe is a moment the shot has to hit partway through. Kling 4.0's model release page lists multi-keyframes as up to 10 keyframe images of input, and Kuaishou says the feature lets a creator set key visual moments and pin character states, scene changes and story beats.

A reference is neither. It is material the model follows, not a frame that appears in the output. MiniMax's docs keep reference generation beside image-to-video as its own mode, where the prompt plus reference images, videos or audio guide a character, a motion, a camera, a style, a voice or an editing rhythm. A first frame pins one instant of a shot; a reference constrains everything the model builds around it.

The distinction matters because prompts are getting crowded. ByteDance's Seedance 2.5 shows the shape in its own published example, a concert scene where the prompt assigns one image to the venue, images six to ten to the orchestra and images eleven to fourteen to the choir. Each numbered image is a reference, and the prompt reads as a cast list.

What "Omni" Adds: Image, Video and Audio in One Request

Image references were already common. Omni reference adds two more channels and lets them run together.

Video references guide motion rather than appearance. Kling 4.0's release notes say video inputs can guide character performance, action, camera movement and narrative pacing, and add that white-model and depth-map videos are accepted, which gives the model a wireframe version of a move to trace.

Audio references guide voice. Kling 4.0's spec table lists reference audio as voice-only, and its Omni Reference description pairs a voice with a subject, so a named character can carry a specified voice. MiniMax H3's reference mode takes audio alongside images and video, with each clip running 2 to 15 seconds.

Seedance 2.5 extends the same idea to size. Its launch post says that "users can now input up to 30 images, 10 video clips, and 10 audio clips as reference materials in a single pass", and that the model preserves the appearances and voices of several characters in shots with multiple subjects.

Holding all three channels at once is the point. Kuaishou describes combining images, videos and subjects in one generation, with a named subject able to carry a specified voice. WaveSpeedAI's AI Cinema Studio exposes the same idea as prompt syntax: upload images, videos and audio clips as references, then point at them in the prompt with an @ handle such as @image1.

Which Models Take References Today

The counts are the meat here, and they swing by an order of magnitude. Each figure below is the number its own maker states on the page linked.

• Kling 4.0 (Kuaishou): up to 15 items across images, videos, voice references and elements in one generation, split as up to 10 images, up to 5 videos with a combined duration up to 30 seconds, and up to 7 subjects including up to 3 video-based. Audio input is voice-only. The launch release states the same ceiling: 15 reference assets and 10 keyframe inputs.
• MiniMax H3: reference mode takes up to 9 images, up to 3 videos of 2 to 15 seconds each, and up to 3 audio clips of 2 to 15 seconds each, with MiniMax's docs stating that "Mixed input is capped at 12 files in total."
• FastH3 Omni (FastVideo at UC San Diego's Hao AI Lab, with Nuva Lab): the release takes up to nine images, three videos and three audio clips, at most 12 references, and outputs 24 FPS video with 32 kHz stereo audio, according to the AICoder release write-up. Nuva Lab's project page describes the surrounding project and notes the weights fall under the MiniMax H3 Community License, which excludes the United States, the European Union, the United Kingdom and South Korea from its default territory.
• Seedance 2.5 (ByteDance): up to 30 images, 10 video clips and 10 audio clips per request, per the maker's own launch post.

Academic work points the same way. UniAVC, a paper hosted by the HKUST Long Group, performs instruction-guided editing across image, video and audio, holding one modality fixed while using it to constrain the other. That is reference guidance applied to editing rather than generation, and it is the clearest statement of where the idea lands: references becoming the way you change a shot, not only make one.

What It Changes in a Real Workflow

The everyday win is consistency inside a scene. A product's color, form and markings are exactly what a text prompt cannot pin down and a reference image can. Kuaishou's release describes creators drawing on the pacing, framing and style of a reference video and replacing the products, characters and assets inside it, which is a shot-for-shot reskin of a proven ad or social clip.

Motion is the second win. A dance, a camera orbit or a fight beat is hard to describe and easy to film once and hand over as a reference video. Kling accepts depth-map and white-model videos for this, and Seedance 2.5's clay render referencing does the same job with a textureless 3D model, fixing a scene's spatial structure, character poses, motion paths and camera angles before a frame is rendered.

Voice is the third. A single audio reference can carry a speaking voice into a scene, which is what turns a reference-heavy request into a short drama or a talking-head ad rather than a silent clip.

The Honest Limits

Three things are worth checking before anyone builds a workflow on references.

First, the numbers do not travel. Kling 4.0 allows 15 assets, MiniMax H3 and FastH3 Omni cap a request at 12 files, and Seedance 2.5 accepts 50. A prompt written for one model will not always fit another, and the ceiling is set per model and per mode.

Second, references can cost more than the output length suggests. The serving platform kie.ai states in its MiniMax H3 reference-to-video documentation that "The total length of reference videos (whole seconds) is billed at the same per-second price and added to the output duration", and that only the first five reference images are free, with each additional image costing extra. Reference audio, it says, is free. That is a serving platform's pricing rather than MiniMax's, but it shows reference input becoming billable input instead of a free upload.

Third, reference-guided output is not character consistency across a film. A reference holds a face, a product or a move inside one generation. Kling 4.0's two-minute extension feature is described as coming soon, and Seedance 2.5 keeps characters stable across several rounds of extension within a session, not across an entire edited production. Consistency over a long film stays a workflow problem, solved by reusing the same reference assets shot after shot and checking each result, not by one setting.

One mechanical catch is worth knowing too. MiniMax's API reference states that image-to-video and reference-to-video cannot be combined: a request that carries reference images, videos or audio cannot also set a first or last frame. WaveSpeedAI's studio takes the opposite approach and treats a first frame as one more asset you can mark. The rule differs by platform, so it pays to check it per mode before assembling a shot.

What Comes Next

The direction is plain. Kling 4.0 shipped Omni Reference into Kling MCP and CLI, so agents can drive it, and Kuaishou says the capabilities will reach enterprise customers and developers through API access. Seedance 2.5's reference budget is the largest of these releases, and the open FastH3 Omni put a 12-file reference set into weights anyone can run locally. Reference ceilings have risen at every one of these releases, and the interface is converging on the same shape: a numbered cast list in the prompt.

To try references today, open Kling AI's creation page, the MiniMax open platform or WaveSpeedAI's AI Cinema Studio, upload your images, video and audio first, then point the prompt at them with a numbered slot or an @ handle.