Nine product angles as references: MiniMax H3 Max on Sume

To make one product video from several photos of the same item, call POST /v1/video-router/generate with model: "minimax-h3-max" and put the photo URLs in reference_image_urls. That model does reference-to-video at 480p, 768p or 1080p for 5 to 15 seconds, per the Video Router page. MiniMax's H3 is described in a third-party summary as taking up to 9 reference images, 3 video clips and 3 audio clips (read 2026-10-05). Treat nine as the vendor figure, and read the limit that Sume enforces from the catalog before you build a nine-photo pipeline.

What nine references are good for

A product has more sides than one photo shows. A reference set lets the model see the front, the back, the label, the cap, and the item in a hand, so the clip does not invent the parts the camera never showed it. Pick the nine by what the clip needs, not by what you have in the folder.

• Two or three full-product views from different angles, on a plain background.
• One close crop of the label or logo, so the text has a source.
• One scale shot, with the item in a hand or on a desk.
• One packaging shot, if the box will appear.
• Skip near-duplicates. Two photos of the same angle add no information.

The request

The body is the same async job envelope as every Video Router model. Send an Idempotency-Key, because a retry after a timeout must not start a second paid job. Every URL must be public HTTPS.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: bottle-angles-001" \
  -d '{
    "model": "minimax-h3-max",
    "prompt": "Slow orbit around the bottle on a marble counter, soft window light",
    "reference_image_urls": [
      "https://example.com/bottle-front.jpg",
      "https://example.com/bottle-back.jpg",
      "https://example.com/bottle-label-closeup.jpg"
    ],
    "resolution": "768p",
    "duration": 8,
    "mode": "async"
  }'

Where the limits are

The docs say that each catalog model can have a different envelope and tell you to read capabilities from GET /v1/video-router/models. Do that once, and let your code read the reference count from the response instead of hard-coding nine. The Video generation page adds that the Seedance 2.x models, Wan 3.0, MiniMax H3 and MiniMax H3 Max accept audio and video references too, and that a model accepts a reference type only if its supported_input_references lists it.

Which MiniMax id for which job

Do not mix the ids up. Only the Max variant is documented for reference-to-video, and Video 1.0 is an alias for Sume Auto, not for H3.

Price and prompt habits

Sume bills list price times 1.25 on every catalog model, so check the price of your resolution before a 100-clip batch. Keep prompts short when you use references: describe the camera and the light, and let the photos carry what the product looks like.

Run the first clip at 480p. A 480p pass shows you whether the references were read correctly: is the label right, is the cap the right shape, does the scale look believable. Fixing a bad reference set at 480p costs less than finding the problem at 1080p. When the clip is right, repeat the same request with a higher resolution and a new Idempotency-Key, because a changed body under an old key returns a conflict, not a new job.

Last, keep the reference photos on stable public URLs. Sume fetches them, and a URL that expires after an hour will fail a retry made the next morning. If the pictures live in private storage, publish them at a public HTTPS address for the length of the job, because the docs reject private-network and non-HTTPS URLs.