Reference-to-video with 3 images, 6 seconds: cost by Sume model

Send three images in input_references, no frame_images, and Sume runs reference-to-video: a 6-second 720p clip is $0.75 on Wan 3.0, $0.75 on Gemini Omni Flash 1.1 and $3.47 on Seedance 2.5. Kling 3 is not an option, because it takes no reference_*_urls.

The request

Each image is a separate input_references entry. Do not add frame_images, which would win and turn the job into image-to-video.

{
  "model": "wan-3.0",
  "prompt": "The three characters walk into a sunlit market",
  "duration": 6,
  "resolution": "720p",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/a.png"}},
    {"type": "image_url", "image_url": {"url": "https://example.com/b.png"}},
    {"type": "image_url", "image_url": {"url": "https://example.com/c.png"}}
  ]
}

Cost by row

Images only, no reference video, so the Seedance reference-video multiplier does not apply. Per-job rounding up to the cent.

Image limits

Omni takes up to 10 images and up to 3 reference clips of at most 3 seconds each, addressed as <IMAGE_REF_0> in the prompt. H3 on Sume has the first five reference images included and charges $0.08 list per image after that, so three images add nothing. The reference limits table has every row.

Which to use

For consistent characters across shots, references are the right mode; a first frame only pins the opening. If three images are all you need and speed of iteration matters, H3 or Wan 3.0 are the cheapest. Seedance 2.5 is the premium row for polish, and the only one that goes to 30 seconds with references among the Seedance family.