TokenLab Adds Kling 3.0 Element References for Video API Workflows

TokenLab now supports Kling 3.0 element references in the video generation API, letting developers anchor specific products, props, or characters to named tags (@name) that stay consistent throughout a generated clip. This closes a common gap in image-conditioned video workflows where a single reference image was not enough to keep multiple subjects visually stable across frames.

Key Takeaways

What Element References Actually Do

Most image-conditioned video generation treats a reference image as a single anchor: you give the model a picture, and it tries to keep the overall look consistent while animating motion around it. That works fine for single-subject shots, but it breaks down fast when a scene needs more than one visually distinct element to persist independently, like a product held by a hand, or two characters exchanging a line of dialogue where each needs its own face and outfit continuity.

Kling 3.0's element references solve this by letting you register multiple named reference images in a single request, then point at them individually from within the prompt text. Instead of one implicit reference, you get explicit, addressable references. The model knows that @shoe refers to reference image one and @model refers to reference image two, and it composes the scene using both anchors simultaneously.

This is a meaningful step up in control for anyone building product video pipelines, character-driven content tools, or ad creative generators where subject consistency across a clip is the difference between usable output and a reshoot.

How to Use Element References in the API

The pattern is straightforward: define your elements, name them, and reference them in the prompt with the @ syntax.

{
  "model": "kling-3.0",
  "prompt": "@shoe rotates slowly on a marble pedestal while @hand reaches in to pick it up",
  "kling_elements": [
    {
      "name": "shoe",
      "image_url": "https://example.com/product-shoe.png"
    },
    {
      "name": "hand",
      "image_url": "https://example.com/hand-reference.png"
    }
  ],
  "duration": 5,
  "aspect_ratio": "16:9"
}

A few practical notes for implementation:

• Element names should be short and unambiguous. Avoid names that overlap with common English words already likely to appear in your prompt text, since that increases the chance of parsing ambiguity.
• Reference image URLs need to be publicly reachable at request time. If your images sit behind an authenticated storage layer, generate a signed or public URL before sending the request.
• You can combine multiple elements in one prompt, but keep the total scene description focused. Piling on more than two or three named elements tends to dilute the model's ability to track each one distinctly, similar to how too many named subjects in a static image prompt reduces per-subject fidelity.
• Test with a short duration first. Element consistency issues, if they occur, show up in the first couple of seconds and are cheaper to catch on a 3-second draft than a full 10-second render.

The One Rule You Cannot Break: No Elements Plus Audio

This is the part worth repeating clearly because it is easy to miss during rapid prototyping: kling_elements and output_audio=true cannot be used in the same request. If you submit both, the request will not process as expected.

If your workflow needs both multi-element visual consistency and generated audio, the current approach is to split the work into two steps: generate the video with element references first, then run an audio generation pass separately and combine the outputs downstream. This is a documented constraint of the current Kling 3.0 integration, not a bug, so build your request validation logic around it rather than treating it as an edge case to catch after the fact.

Element References vs. Broader Reference-to-Video Workflows

Element references are one tool in a growing set of reference-to-video capabilities available through TokenLab's video API. It helps to know when to reach for which:

Workflow	Best for	Reference count	Notes
Single image-to-video	Simple animation of one static image	1	Works across most supported video models, including Seedance and PixVerse V6
Kling 3.0 element references	Multi-subject scenes needing independent consistency	2-3 named elements	No audio in the same request
Style or motion reference	Applying a visual style or camera motion pattern	1 style reference + prompt	Available on select models, check per-model docs
Text-only prompting	Fast iteration, no visual anchor needed	0	Fastest to prototype, least controllable

If you are building a product demo generator, element references are usually the right call. If you are doing simple animation of a single hero image, plain image-to-video is faster and cheaper to iterate on. For teams comparing video models more broadly, including how Kling 3.0 stacks up against Veo 3 and other options for different use cases, the best AI video models for API use in 2026 breakdown is a good starting point.

Implementation Checklist

Before shipping a Kling 3.0 element-reference workflow to production, confirm the following:

• Each element has a unique, unambiguous name
• Reference image URLs are publicly accessible and stable for the duration of processing
• Prompt text correctly tags each element with @name syntax
• output_audio is not set to true when kling_elements is present
• Request validation catches the audio-plus-elements conflict before it reaches the API
• Test renders use short durations before committing to full-length generation
• Total named elements per request stays at two or three for best consistency

FAQ

Can I use more than two element references in a single Kling 3.0 request? Yes, the API does not hard-cap the count, but practical consistency tends to degrade as you add more named elements to a single scene. Two to three is a reasonable working limit for most product and character use cases.

What happens if I send both kling_elements and output_audio=true? The request will not process correctly since these two parameters are mutually exclusive in the current Kling 3.0 integration. Validate this combination client-side before sending the request to avoid wasted calls.

Is element reference support specific to Kling 3.0, or available on other models too? Named element references with @name tagging are specific to Kling 3.0 in the current API. Other supported video models have their own reference-to-video patterns, typically limited to a single reference image per request, so check the model-specific docs before assuming feature parity.

Sources and Freshness

This article reflects the TokenLab video API documentation and Kling 3.0 integration behavior observed on 2026-07-07. For the current parameter reference, see the create video API reference and the video generation guide. API behavior can change, so always check the live docs before finalizing a production integration.

---

Ready to add Kling 3.0 element references to your video pipeline? Get your TokenLab API key and check the video generation guide to start building multi-subject video workflows today.

Related Reading and Next Step

Element references expand what's possible with Kling 3.0, but choosing the right video model and understanding costs still matters before you build a production workflow. If you're comparing options, the Best AI Video Models API Guide: How Developers Should Choose Video Generation Models walks through the tradeoffs across providers. For a closer look at Kling specifically, the Kling AI API Pricing Guide: Cost, Workflow, and Alternatives breaks down pricing and workflow considerations. And if you're weighing alternatives, the Seedance API Guide: When to Use It for AI Video Generation covers when that model fits better.

Model capabilities and pricing change frequently, so verify current model versions and rates directly before relying on them for high-volume production use. When you're ready to start testing element references or other video workflows, Create an API key and begin building.