Animate an infographic with Wan 3.0: first frame or reference image?

To animate an infographic with Wan 3.0 and keep its layout, send it as a first frame (frame_images with frame_type: "first_frame" on /v1/videos, or image_url on the Video Router). To get new motion that only borrows its look, send it as a reference (input_references or reference_image_urls). Sume's Video generation docs define the two modes: frame images start an image-to-video job, references give the model visual guidance but not exact frames, and if you send both fields, frame_images controls the mode. Alibaba's Wan 3.0 README (read 2026-10-05) lists up to 20 reference assets and text rendering in 12 languages; read both as upper bounds, not promises for one chart.

What each mode does to an infographic

As a first frame, the infographic is frame zero: the headline, icons and numbers are exactly what you drew. The model then animates forward, and over a few seconds the layout can melt, because nothing holds the text in place after the first frame. As a reference, the infographic is only guidance, so the clip will not contain your layout at all, only a similar palette and style.

Keep the clip short

The longer the clip, the more the text can drift. For a layout-faithful animation, ask for a short duration, such as 4 to 6 seconds, and use a first frame plus a last frame designed in your own tool, so the model only has to move between two states you made. The docs list end_frame support for Wan 3.0 in the catalog, and the Video Router's image-to-video fields are image_url and end_image_url.

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: infographic-001" \
  -d '{
    "model": "wan-3.0",
    "prompt": "Bars grow upward one by one, icons pulse once, camera static.",
    "image_url": "https://example.com/infographic-start.png",
    "end_image_url": "https://example.com/infographic-end.png",
    "resolution": "720p",
    "duration": 5,
    "mode": "async"
  }'

When the text must be exact

Do not trust motion with small type. Animate the bars, icons and background, then lay the real labels over the result. Timeline compose does one still plus one video, and video captions burns authored cues. Put big numbers in captions, keep legends in the still, and treat anything the model draws as decoration.

Keep the prompt short and about motion only, since the picture already says what the content is. A practical order of work: design the start and end frames as ordinary images in your design tool, export them at the same size and the same aspect ratio as the output you will request, host them at stable https URLs, and generate a 480p draft first. The draft shows you whether the in-between motion is acceptable before you pay for the larger size. If the motion keeps breaking the layout, split the animation into two five-second clips, each between its own pair of frames, and join them.

Check the last frame with a still before you ship. If the last frame no longer matches the data, shorten the clip or replace the model's frames with your own final card using Timeline 1.0.