Kling 3 on Sume: text or start and end frames, no reference inputs

Sume's kling-3 row takes a text prompt, or a start frame and an end frame, for 4 to 15 seconds at 1080p, and it accepts no reference images, reference video or reference audio. If your Kling workflow depends on character references, use Wan 3.0, a Seedance row or Gemini Omni Flash 1.1 on Sume instead.

The limits below come from Sume's Video Router capability table and request schema, and the prices from its pricing code.

What the row accepts

The capability entry for kling-3 (named Kling Video v3 Pro) sets text-to-video, image-to-video and end frame to true, and reference images, reference video and reference audio to false. The constraint text says there are no reference_*_urls. The row renders at 1080p only; a 720p request from older callers is accepted but still renders 1080p. Aspect ratios are 16:9, 9:16 and 1:1.

Native audio is available, and it changes the price, as the table shows.

Billed cost with and without audio

Sume bills the fal list price times 1.25. The list is $0.168 per second with audio and $0.112 without, so Sume's billed rates are $0.21 and $0.14 per second.

Because the resolution is fixed at 1080p, the only price levers are duration and the audio toggle.

What to use when you need references

Per the Video Generation docs and the request schema, rows that take reference images are Seedance 2.x (9 images), Wan 3.0 (10), MiniMax H3 (9) and Omni (10). A 5-second Wan 3.0 clip at 1080p bills $1.25 on Sume, close to Kling's $1.05 with audio for the same length, and it accepts references.

If the Kling look matters to you, there is no stand-in; you would keep that work outside Sume. If consistency of a character matters more than the look, a reference-capable row is the practical choice.

• Start and end frame control: Kling 3, Wan 3.0 or Seedance.
• Character references: Wan 3.0, Seedance 2.5 or Omni.
• Square output with audio: Kling 3 or Wan 3.0.

Two practical notes

First, the end frame is the trick that makes Kling 3 useful without references: generate or pick a first and last still, and let the model fill the motion between them. That gives you tight control of where a clip ends, which a reference image does not.

Second, do not send bitrate_mode or reference fields to this row. Validation rejects them before any money is reserved. Check supported_frame_images and supported_input_references in the live catalog, and keep an Idempotency-Key on the submit, as the Video Generation docs recommend.

Planning a clip around two frames

Start and end frames suit before-and-after shots: a product closed then open, a room empty then furnished, a face neutral then smiling. Produce both stills in the same style and ratio, with the same lighting, so the model has to invent only the motion. The row accepts 16:9, 9:16 and 1:1, so crop the stills to one of those before you send them.

A 10-second clip with audio bills $2.10, so test the framing with a 4-second clip first, which is the shortest the row accepts. That is $0.84 with audio on and $0.56 with it off, and the stills you reuse cost nothing extra.