Best AI video model for product videos: pick the row by input

There is no single best model for product videos on Sume; the right row follows from what you can show it. If you have one clean packshot, use a row that takes a first frame (frame_images). If you have several angles of the same product, use a row that takes reference images (input_references). If you have both a start and an end state, use a row that takes an end frame. The catalog on origin/main lists which rows do which, and the request is rejected with a 400 if you send an input the row does not list, so you find out before you are billed.

This post ranks nothing by looks, because the Sume docs publish no quality benchmark and this post invents none. It sorts rows by the inputs and limits that decide whether a product job can run at all.

Which rows take a product photo as input?

The table lists what the catalog says each row accepts. Mode is inferred from the fields you send: frame_images makes image-to-video, input_references makes reference-to-video, and if both are present frame_images wins (Generate video).

How do I choose between a first frame and references?

A first frame pins the opening picture, so the product looks like your photo at second zero and the model animates away from it. References are looser: the model sees the product as an idea to place in a new scene, which suits a lifestyle shot where the product is not the first thing on screen. Use references when you have more than one angle, because a first frame can only hold one.

Two limits matter for product work. Kling 3 on Sume takes no reference images, so its product jobs start from a first frame. Gemini Omni Flash 1.1 caps a reference video at 3 seconds each and tags references in the prompt as <IMAGE_REF_0>, counted from zero in list order.

• One angle, exact packaging: first frame on wan-3.0, kling-3, seedance-2.5 or gemini-omni-flash-1.1.
• Three or more angles: references on wan-3.0 (up to 10 images) or gemini-omni-flash-1.1 (up to 10 images).
• Product reveal from closed to open: first frame plus end frame.
• Only a still and no budget for experiments: grok-imagine-video-1.5 animates one image, 4-15 s, up to 720p.

What about length, ratio and sound?

Most product ads are 6 to 15 seconds, which every row above covers. Vertical 9:16 is listed by every row except grok-imagine-video-1.5, which takes no aspect_ratio and follows your photo. gemini-omni-flash-1.1 allows only 16:9 or 9:16, so a square 1:1 cut needs another row. Audio is native on most rows (not Grok); minimax-h3, minimax-h3-max and gemini-omni-flash-1.1 always produce it, with no toggle.

How do I test three rows on one photo cheaply?

Run the same first frame through three pinned ids at the lowest resolution each row lists, compare, then rerun the winner at delivery resolution. Pin the ids instead of sending sume/auto, because Auto does not tell you which family served the job. The docs list per-model prices in the catalog (pricing_skus), so read those before a batch.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "wan-3.0",
    "prompt": "Slow orbit around the bottle on a marble counter, soft window light",
    "duration": 6,
    "resolution": "480p",
    "aspect_ratio": "9:16",
    "frame_images": [{
      "type": "image_url",
      "image_url": {"url": "https://example.com/packshot.jpg"},
      "frame_type": "first_frame"
    }]
  }'