Kling v3 Pro multi_prompt shots on fal vs Sume's one clip per shot

If you want Kling v3 Pro to plan several shots inside one video, fal's API page lists a multi_prompt field for that. Sume's kling-3 does not pass it: the request Sume builds carries one prompt, a duration, an audio flag and the start and optional end image. On Sume you make one clip per shot and join them with Timeline 1.0.

Which is better depends on whether you value one continuous generation or the ability to regenerate a single shot. This post sets out what each side documents so you can choose.

What does fal's Kling v3 Pro page list?

The fal API page for Kling Video v3 Pro image-to-video lists these inputs: start_image_url (required), optional end_image_url, duration from 3 to 15 seconds (default 5), generate_audio (default true), and shot_type with the values customize or intelligent (default customize). It also lists multi_prompt, described as a list of prompts for multi-shot generation that divides the video into multiple shots, each shot with its own duration of 1 to 15 seconds.

The page does not state a price or an aspect-ratio list in the part I could read, so I make no claim about either for fal.

What does Sume's kling-3 send?

In Sume's provider code, the Kling v3 Pro payload is built from prompt, a rounded duration clamped to 3 to 15, generate_audio (default true), a fixed negative prompt, cfg_scale, and for image-to-video the start_image_url plus an optional end_image_url. There is no multi_prompt or shot_type in it. A text-to-video request adds an aspect ratio of 16:9, 9:16 or 1:1.

The public contract is narrower still: provider.options is rejected with a 400 on POST /v1/videos because allowed_passthrough_parameters is empty for every model, so you cannot smuggle multi_prompt in. In the Video Router catalog, kling-3 lists text-to-video, image-to-video and end frame, with 720p and 1080p, 4 to 15 seconds, and 16:9, 9:16 or 1:1.

How do the two approaches compare?

Note the minimum: Sume's catalog floor for kling-3 is 4 seconds even though the provider code clamps at 3, so request 4 or more.

How do I do multi-shot on Sume?

Write the shot list yourself: for each shot, a still (first frame), a motion prompt and a length. Submit each as its own POST /v1/videos with frame_images, an Idempotency-Key such as spot-7-shot-03, and model: "kling-3". Then place the clips in video[] on a timeline with declared start values.

What you lose is the model's own idea of how a scene should be cut. What you gain is control: a bad shot 3 costs one clip, not the whole sequence, and you can mix models, using a different catalog model for a shot that needs references or a longer length.

{
  "model": "kling-3",
  "prompt": "Slow push-in on the red sneaker, soft studio light",
  "duration": 5,
  "resolution": "1080p",
  "generate_audio": false,
  "frame_images": [
    {
      "type": "image_url",
      "image_url": { "url": "https://media.sume.com/artifacts/artf_demo/shot-03.png" },
      "frame_type": "first_frame"
    }
  ]
}

When should I still prefer multi_prompt?

If you are already on fal with Kling and your shots are short, tightly coupled beats that you want planned as one, multi_prompt is the more direct tool and Sume does not replace it. The docs I read do not say how well a model keeps a character stable across the shots it plans, on either side, so test with a shot list you care about.

Where Sume fits is when the film has to be edited after generation, billed per clip, and rebuilt from parts. Per-shot jobs, a declared timeline and unbilled frame checks make that loop short.