MiniMax H3 takes 9 image references: a slot plan for one product clip

Nine reference images are enough to describe one product from every side, plus the place and the look, but only if each slot has a job. A third-party summary (the Hugging Face blog on MiniMax H3, read 2026-10-05, not the vendor page) lists up to 9 reference images, 3 video clips and 3 audio clips for H3, which launched on 2026-07-31. Sume's Videos docs say MiniMax H3 and H3 Max accept image, audio and video references, and tell you to read each model's limits from the catalog.

A 9-slot plan

Fill slots in an order you can name in the prompt. The list is a working plan, not a MiniMax rule:

Which slots to drop first

Nine is a ceiling, not a target. If the clip is a simple push-in on a product, slots 1, 5 and 8 may be enough. Drop in this order: the style frame (9), the empty setting (7), the top view (4), the back view (3). Keep the identity image, the logo close-up and the light reference as long as you can.

Each added reference is also another URL that must be reachable. The docs say Sume rejects localhost, private-network and non-HTTPS URLs for media inputs, so host every image on a public HTTPS address before you submit. A single unreachable URL fails the whole request.

Draft, then final

Run the first attempt at 480p with the three slots above, check that the product looks like the product, and only then add slots or raise the resolution. The docs show 480p, 768p and 1080p on H3 Max, with 1080p a latent refinement from native 768p. A 10-second 480p draft costs less than a 10-second 1080p final under the per-second billing, so test the plan cheap and finish expensive.

Sending it

Video Router takes URLs in reference_image_urls. Per the Video Router docs, minimax-h3-max does reference-to-video through the reference_*_urls fields at 480p, 768p or 1080p for 5 to 15 seconds. The example sends three of the nine slots; the rest follow the same shape. Check your id against GET /v1/video-router/models first, because each model can have its own envelope.

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: h3-product-001" \
  -d '{
    "model": "minimax-h3-max",
    "prompt": "Slow orbit around the product from image 1, ending on the logo close-up from image 5. Soft window light as in image 8.",
    "resolution": "768p",
    "duration": 8,
    "reference_image_urls": [
      "https://example.com/front.png",
      "https://example.com/logo.png",
      "https://example.com/light.png"
    ],
    "mode": "async"
  }'

Tips

• Name each image by its slot and role in the prompt. The pages read give no H3-specific tag syntax, so test whether plain words work before you rely on them.
• More references do not mean more control. Two clean views beat six cluttered ones.
• Use the same product, same lighting and same scale across slots 1 to 6, or the clip will blend the differences.
• Draft at 480p. The docs show H3 Max at 480p, 768p and 1080p, where 1080p is a latent refinement of native 768p.
• Billing is the provider list times 1.25 per output second, and the docs say jobs reserve at submit. Read usage.cost for the billed amount.