Wan 3.0's 20 reference assets: what an ad request can carry on Sume

Alibaba's Wan 3.0 README says the model takes up to 20 reference assets, and that they can be documents, webpages, text or images (read 2026-10-05). On Sume, the wan-3.0 Video Router model accepts 2 to 30 seconds, and the Video generation page says it accepts image, audio and video references. It does not say that Sume forwards documents or webpages. So for an ad brief, send the images as references, and put the document and webpage content into the prompt as text.

Vendor facts versus what Sume sends

Keep two columns in your head. The vendor column says what the model can do. The Sume column says what a request to Sume can carry. Only the second one is a promise to your integration.

A brief that fits the prompt

An ad brief usually starts as a product page and a PDF. Reduce it before you send it. Pull the three claims you will show, the product name exactly as it is printed, and the offer. Write them as short lines in the prompt, and send the photos as references. This is cheaper and more predictable than hoping a model reads a long document, and it keeps every sentence in the clip something you chose.

Wan 3.0 also renders text in 12 languages, per the same README. If the clip must show a line on screen, put the exact string in the prompt in quotes.

Request shape

Use the explicit Video Router so the model id is pinned. Duration must be inside 2 to 30 seconds.

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: wan-ad-brief-001" \
  -d '{
    "model": "wan-3.0",
    "prompt": "Three shots of the kettle. Shot 1 on-screen text: \"Boils in 90 seconds\". Shot 2: pour into a mug. Shot 3: logo end card.",
    "reference_image_urls": ["https://example.com/kettle-front.jpg", "https://example.com/kettle-side.jpg"],
    "duration": 15,
    "aspect_ratio": "9:16",
    "mode": "async"
  }'

Check before you scale

Call GET /v1/video-router/models/wan-3.0 and read capabilities and supported_input_references. If the count of references matters to your design, take it from that response. The vendor's number of 20 is a ceiling for the model, and a hosted route can set a lower one. Test one clip with your real reference set before you queue a batch, and join clips with Timeline when you need a story longer than one generation.

When 20 is the wrong target

More references do not make a better ad. Each added image is one more thing the model has to reconcile, and a catalog photo with a busy background can pull the whole clip toward that background. For a single product, three to six clean images usually cover what the model needs. Use the wider set only when the ad has several distinct items, such as a kit with a box, a bottle and a tool.

A second use for the large budget is variants. Keep the product references fixed, and swap one reference for the setting: a kitchen, a gym bag, a desk. You get a family of ads that share the same product rendering and differ in context. Run them as separate requests with separate idempotency keys, and compare the clips side by side before you pick one.

If the ad needs more than 30 seconds, do not stretch one request. Generate two clips from the same reference set and join them in Timeline, so each half has its own prompt and its own on-screen line.