Lock the look with one image, then animate it on three video models

To lock the look, make one image you approve, and pass it as the first_frame in frame_images to each video model you want to try. All three clips start on the same picture, so the character, the clothes and the set match, and only the motion differs. On Sume that is wan-3.0, kling-3 and minimax-h3-max through the same POST /v1/videos body with a different model value.

The same trick works in the other direction. Once you have chosen a model, keep the image and make several clips from it with different prompts, and the whole set matches. That is how a short series with a recurring character is built from one approved picture instead of from luck.

This turns model choice into a fair test. If each model started from its own text prompt, you would be comparing both the model and its interpretation of the words.

Step one: the still

Make the image with the image API at POST /v1/images, with a public model id such as google/nano-banana-2.1, or use a photo you own. Check it at full size for faces, hands and any text, because every frame of every clip inherits what is wrong in it. Pick a framing with headroom, since the models will move the camera or the subject and a tight crop leaves nothing to move into.

Match the image to the aspect ratio you will request. Gemini Omni Flash 1.1 takes only 16:9 and 9:16, and other models list their own ratios in supported_aspect_ratios, so choose a ratio that all your candidate models accept.

Step two: one body, three model ids

Run the loop below with an Idempotency-Key per model, so that you can retry one model without creating a duplicate of another. Each call returns an id; poll each one, or add a callback_url to the body, and download the three files when they finish. Keep the file names tied to the model id so the comparison stays honest when you review them a day later.

Use the same prompt, the same duration and the same first frame. Change only model and, if needed, the resolution. The duration must be inside every model's window, which for these three means 5 to 15 seconds.

for m in wan-3.0 kling-3 minimax-h3-max; do
  curl -s -X POST "https://api.sume.com/v1/videos" \
    -H "Authorization: Bearer $SUME_API_KEY" \
    -H "Idempotency-Key: look-lock-$m-v1" \
    -H "Content-Type: application/json" \
    -d '{
      "model": "'"$m"'",
      "prompt": "Slow push-in, she turns toward the window",
      "duration": 6,
      "resolution": "720p",
      "frame_images": [{
        "type": "image_url", "frame_type": "first_frame",
        "image_url": {"url": "https://example.com/approved.png"}
      }]
    }'
done

What differs between the three

The catalog rows give the facts that matter for the comparison. Note that H3 Max resolution is 768p and not 720p, so request the nearest tier it lists.

Reading the three results

Resist the urge to change the prompt between runs. If the prompt differs, the comparison is no longer about the model. Change one variable at a time: first the model, then, for the one you keep, the prompt, then the length. A table of your own notes with one row per run is enough, and it makes the choice easy to explain to a client or a teammate who was not there.

Compare the clips on what the image fixed and what it did not. Check whether the face holds for the whole clip, whether the clothing and the set stay the same, and whether the motion matches the prompt. Then compare what the models added: sound, camera movement and speed. Pick the one whose motion you want and re-run that model for the final length. If the clips drift from the image, shorten the clip or simplify the prompt, since long clips give a model more time to move away from its first frame.

For the full list of input limits per model, see the reference limits table.