Plant shop care-tip loop: Seedance 2.5 with one image as both frames

Send one photo as first_frame and last_frame to Seedance 2.5 on Sume for an 8-second 1:1 loop-style clip: the request body, $4.63 at 720p, and the limits.

5 min readSume
All posts

For a loop-style clip, send the same image as both first_frame and last_frame to seedance-2.5 and describe a small motion in the prompt. On Sume an 8-second 1:1 clip at 720p (960x960) costs $4.63. The model starts and ends on your photo, so the end of the clip sits on the start and a short loop in a social feed or a shop page hides most of the seam.

That is the whole trick. It is not a guarantee of a seamless loop: motion in the middle can differ in speed or direction from what the first frame implies, and the cut from the last frame back to the first is only invisible if the motion eases to rest. Test one take before you buy a set.

The request

Sume's /v1/videos takes frame_images entries with frame_type set to first_frame or last_frame, as listed in the model's supported_frame_images (see the video docs, read 2026-10-07). Seedance 2.5 reports both. URLs must be public HTTPS.

{
  "model": "seedance-2.5",
  "prompt": "A potted monstera on a sunlit windowsill, leaves swaying gently, the camera holds still",
  "duration": 8,
  "resolution": "720p",
  "aspect_ratio": "1:1",
  "frame_images": [
    {"type": "image_url", "image_url": {"url": "https://example.com/monstera.png"}, "frame_type": "first_frame"},
    {"type": "image_url", "image_url": {"url": "https://example.com/monstera.png"}, "frame_type": "last_frame"}
  ]
}

Price and mode

Eight seconds at 960x960 is 172,800 video tokens. At 480p (640x640) the same clip is $2.06, the cheapest ratio and resolution in the family for a square clip, and enough to check whether the idea reads. For a 12-plant catalog, 12 clips at 720p is $55.56.

A reference-to-video request (input_references) is a different mode: it conditions the look but does not pin frames. If you send frame_images and input_references together, Sume treats the request as image-to-video, per its docs.

What works and what to skip

Use a still with a clean subject and a quiet background; a busy photo gives the model more to drift. Ask for gentle motion (leaves, steam, a curtain) and say that the camera holds still. Avoid real people in the frame images; Seedance policy on real faces is covered in another post.

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume