Kling 3.0 element references are not on Sume's kling-3: use these

Kling's VIDEO 3.0 guide lists input types that include start frame plus element reference and multi-image or video references as elements (read 2026-10-04). That is how you keep one character or product consistent across shots.

On Sume, kling-3 does not take them. The catalog capabilities for the row say reference_images, reference_videos and reference_audios are all false, and the constraint reads "no reference_*_urls". Reference inputs sent to it are refused, so use it for text-to-video or start and end frames only. The video generation docs show where reference inputs go for the models that take them.

Models that do accept references

Read the supported_input_references field of each row at GET /v1/videos/models rather than guessing. From the registry as of this post:

A workable route for one consistent character

If your Short needs the same face in every shot and you like Kling's motion, generate a still of the character first, then pass it as the start frame to each kling-3 clip through frame_images with frame_type: first_frame. A start frame fixes the opening, not the whole shot, so expect drift later in the clip. When consistency matters more than the Kling look, switch the clip to wan-3.0 or seedance-2 and send the same reference images with every request.

Whichever model you pick, keep the price in view. Kling v3 Pro bills $0.14 per second silent or $0.21 with audio at Sume's list times 1.25; Wan 3.0 at 720p bills $0.125 per second.