Six-shot character consistency: Kling vs Sume references

To keep one character across several shots on Sume, give every shot request the same reference image, or create an Avatar once and reuse its handle in each avatar video. The docs describe reference images as visual guidance, not a guarantee, so review each shot.

What the Kling guide says

A Luma Labs guide to Kling describes 15-second clips with native audio, 2 to 3 minutes of generation per clip, and character consistency across up to 6 shots. It adds that standard exports lacked HDR and EXR as of early 2026. Those are the guide's statements, read on 2026-10-03.

Option 1: the same reference in every shot

On POST /v1/videos, input_references supplies reference images for reference-to-video. The model uses them as visual guidance rather than exact frames, and only models whose supported_input_references lists a type accept it. The legacy Video 1.0 URL takes reference_image_urls (1 to 9 images). Send the same character images with each shot's prompt, and keep the clothing and setting in the prompt text.

The docs state no shot count limit and make no consistency guarantee, so do not assume parity with the six-shot figure above.

Option 2: an Avatar handle

Avatar 1.0 creates a reusable avatar from a prompt, structured traits or a reference photo, stored under an avatar_handle. Each POST /v1/avatar-1.0/talking-video request references that handle, with scripts of 4 to 60 seconds. This suits a presenter who speaks in every shot; one resolved avatar is supported per final video.

Assemble the shots

Generate each shot as its own job, then join them with Timeline 1.0, which takes up to 200 ordered video[] slots over one audio spine and returns a single MP4. Store every job id; a failed shot can then be regenerated alone without resubmitting the rest. See Jobs and results.