Wan 3.0 'Image 1' and 'Audio 1' prompt labels on Sume

Alibaba's Wan 3.0 reference page tells you to point at media in the prompt as "Image 1", "Image 2", "Video 1" and "Audio 1", by position in the media array. Sume's docs do not define a label syntax for wan-3.0 or say how labels are read, so try the same labels and check the result at 480p before a long run.

What does Alibaba say about referring to media?

The Wan3.0 API reference (read 2026-10-09) lists these media types and the labels used in prompts.

How does Sume order the references?

On POST /v1/videos you send one input_references array that mixes types, up to 12 items. Sume's docs do not say how entries are numbered for labels, so keep each type in list order and number your labels the same way, then check the result.

If you send frame_images as well, the request is treated as image-to-video and input_references are ignored. Keep the two modes apart.

What does a labeled request look like?

This body uses two images and one audio clip. The prompt names each one.

{
  "model": "wan-3.0",
  "prompt": "Image 1 is the barista, Image 2 is the shop. Audio 1 is her voice: she greets a customer at the counter.",
  "duration": 5,
  "resolution": "480p",
  "aspect_ratio": "16:9",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/barista.png"}},
    {"type": "image_url", "image_url": {"url": "https://example.com/shop.png"}},
    {"type": "audio_url", "audio_url": {"url": "https://example.com/voice.mp3"}}
  ]
}

What does the test cost?

fal lists Wan 3.0 at $0.05 a second at 480p (fal, read 2026-10-09) and Sume bills list times 1.25, so a 5-second test is 5 x $0.0625 = $0.3125. Run it first, look at whether the right face ended up on the right person, and only then move to 720p. If the labels were ignored, describe each reference by what it shows instead of its number.