'kling-3 supports text-to-video or start/end frames only': the fix

The Video Router message "model kling-3 supports text-to-video or start/end frames only (no reference_*_urls)." means the request contained reference_image_urls, reference_video_urls or reference_audio_urls. For kling-3, send a prompt alone, or a prompt with image_url and optionally end_image_url. For reference-guided shots, use a model with reference support.

What kling-3 does take

In the catalog kling-3 (Kling Video v3 Pro) has text-to-video, image-to-video and an end frame, no reference images, videos or audios, resolutions of 720p and 1080p at the catalog level, and 4 to 15 seconds. Its catalog entry also carries a separate audio-on list price. One caveat: the Video Router validation lists the models that may use 1080p, and kling-3 is not among them, so check the live model endpoint for what the route accepts before you pin 1080p.

Convert references into frames

If you can pick one image as the opening, a reference image becomes a first frame, and the request is valid for Kling. The converter does exactly that, refusing a request that has anything but images, since videos and audio have no frame equivalent. It prints the body and sends nothing.

def kling_ready(body: dict) -> dict:
    out = dict(body)
    if out.get("reference_video_urls") or out.get("reference_audio_urls"):
        raise ValueError("kling-3 has no video or audio references; change model")
    refs = out.pop("reference_image_urls", [])
    if refs:
        out.setdefault("image_url", refs[0])
    return out

print(kling_ready({
    "model": "kling-3",
    "prompt": "The courier smiles and waves",
    "reference_image_urls": ["https://example.com/courier.png"],
}))

When the swap is wrong

A reference image and a first frame are not the same instruction. A reference tells the model what a subject looks like; a first frame is the picture the clip starts from. If you only need the look of a character across many shots, a reference model is the right fit, and forcing it into a first frame will fix the error while changing the shot. The related post on kling-3 and motion_video_url covers the other route that takes a clip.

Before you migrate a whole workflow, run one reference-style shot on both models at the same duration and resolution. Compare how well the subject holds, and compare the per-second price in each catalog entry. The cheapest model that keeps the character recognisable is the right one; a model that is cheaper per second but needs three retries is not.

Idempotency keys also deserve a thought. A corrected request is a new body, so give it a new Idempotency-Key, or a replay of the earlier key may return the original outcome instead of running your fix.