Vidu Q4, Wan 3.0 and Seedance 2.5: reference inputs compared

Vidu Q4 Preview takes up to 15 reference images and up to 3 MP3 audio references, with no video references on its reference-to-video page. Wan 3.0 takes 10 images, 5 videos and 5 audio clips, 20 items in all. On Sume, wan-3.0 keeps Alibaba's 10/5/5 caps and seedance-2.5 accepts 12 references in total (at most 9 images, 3 videos and 3 audio files). Vidu Q4 itself is not listed on Sume.

How do the reference limits line up?

Counts first, then what each model's page says about prompts.

Where does each model fit a reference-heavy job?

Pick by the mix of media you hold.

• Many faces or products and no audio: Vidu's 15 images are the highest count here, but you cannot send them to Sume.
• Reference clips and audio together: wan-3.0 is the row with explicit per-type caps and totals.
• A mixed set up to 12 files: seedance-2.5 caps images at 9, videos at 3 and audio at 3, with 12 in total.
• Clips up to 30 seconds: both Sume rows go to 30 s; Vidu Q4 Preview stops at 16.

How do you check the live limits?

Read supported_input_references for each id from GET /v1/videos/models. The video generation docs say the Seedance 2.x models, Wan 3.0 and MiniMax H3 accept audio and video references, while Gemini Omni Flash 1.1 accepts video but not audio. The counts are enforced at submit and return a 400 naming the model and the field.