AI video from multiple images: frames, references, stills

You can make one AI video from multiple images, but a single generation on Sume pins at most two of them as frames: one as the first frame and one as the last. Two images make one transition clip. With more images, you have three options: generate one clip per neighboring pair and join them, send the images as references that guide a single clip, or hold them as stills in an edited sequence.

The fields and limits come from Sume's Video generation, Video Router, and Timeline 1.0 docs, read on 2026-09-27. Anything described as current behavior is read from Sume's API code.

How do I make an AI video from two images?

Send both images in frame_images on POST /v1/videos: one entry with frame_type: "first_frame" and one with "last_frame". That is image-to-video, and the prompt describes what happens in between. The array takes at most two entries, a last frame needs a first frame in the same request, and the model must list last_frame in supported_frame_images on GET /v1/videos/models. Image-to-video API: first and last frame lists which models take each frame.

{
  "model": "seedance-2",
  "prompt": "Slow push-in as the street goes from morning to night and the shop lights come on",
  "frame_images": [
    {
      "type": "image_url",
      "image_url": { "url": "https://example.com/street-day.png" },
      "frame_type": "first_frame"
    },
    {
      "type": "image_url",
      "image_url": { "url": "https://example.com/street-night.png" },
      "frame_type": "last_frame"
    }
  ],
  "duration": 8
}

How do I turn more than two images into one video?

Pick the approach by what the images should do. To pass through every image in order, chain pairs: image 1 to 2, then 2 to 3, and so on, so each clip ends on the image the next one starts from. Four images make three clips, and each clip is its own job and its own bill. Then join the clips; Timeline 1.0 does it in one render from their Sume-hosted URLs.

Can one AI clip use all my images at once?

Yes, as references: send each image as an image_url entry in input_references for reference-to-video. The docs describe references as style or content guidance that the model uses rather than exact frames, so nothing promises that every image appears in the clip, or in any order. In the current catalog, kling-3 and grok-imagine-video-1.5 take no references.

Gemini Omni Flash 1.1 lets the prompt point at a reference by position: <IMAGE_REF_0> is the first image, counted from 0 in list order, as covered in Gemini Omni Flash 1.1 video API. References and frames do not mix: if a request carries both, frame_images takes precedence and the request is image-to-video.

What if the photos should just play in order?

If the images are already on Sume, such as earlier Sume outputs, nothing needs generating. Timeline 1.0 takes a still in a video slot as a static hold, so each image stays on screen for its slot's duration, with an optional fade, wipe, slide, or dissolve of up to 1 second between slots. It needs one audio spine or audio.mode: "silence". Your own photo files cannot go in directly: every slot must be this workspace's media.sume.com file, and the public API documents no upload route. Image slideshow video API has the full request.

What are the limits?

• Two frames per generation, first and last, and frame URLs must be public HTTPS.
• Each clip runs 2–30 seconds depending on the model; GET /v1/videos/models lists each model's supported_durations.
• Chaining N images costs N − 1 generations; the docs make no promise about how the motion between two frames looks.
• Reference caps differ by model, and in current code a request over a cap is refused with 400 unsupported_capability.
• Timeline takes only Sume-hosted media, 1–200 slots per render.