Wan 3.0 document-to-video claim vs what Sume accepts as reference

Alibaba says Wan 3.0 reads PDFs, slides and docs. Sume's catalog lists image, video and audio references for it, not documents. Read the field before you plan.

5 min readSume
All posts

Do not plan a Sume pipeline around feeding Wan 3.0 a PDF. Alibaba's launch post says Wan3.0 accepts documents (PPT, PDF, DOC, XLS and more), one file or link of up to 100 MB and 50 pages. The Sume docs describe wan-3.0 as taking 2 to 30 seconds and accepting audio and video references, and the reference types the docs name are image, video and audio. No document type is listed.

That is the useful distinction: a vendor capability on a vendor's own cloud is not the same as a field on Sume's /v1/videos. The Sume docs say a model accepts a reference type only if supported_input_references includes it.

Vendor statement and Sume field

Wan 3.0 inputs: Alibaba's launch post and Sume's documented fields (read 2026-10-07)
InputAlibaba saysSume docs
Text promptYesprompt
ImageYesframe_images, input_references
VideoYesinput_references (video)
AudioYesinput_references (audio)
DocumentYes, 1 file or link, 100 MB, 50 pagesNot listed
Output tiers480P, 720P, 1080P480p, 720p, 1080p
Max length30 seconds2 to 30 seconds

How to work around it

If your Sora workflow began with a script or a slide deck, do the conversion yourself before the video call. Extract the text and write the prompt, or export slides as images and pass them as references where the model accepts images. The conversion step is yours, and it keeps the request inside fields the catalog declares.

Check before you build. Call GET /v1/videos/models, find wan-3.0, and read supported_input_references. If a document type ever appears there, the field list is the proof, and until then treat the vendor claim as a vendor claim.

Why this is worth a post

A migration is the moment buyers compare spec sheets. Launch posts list every capability the vendor's own platform exposes, and a router exposes the subset it has wired. The difference costs a week when it is found after the design is done, and nothing when it is found in the catalog first.

Fields that do exist

  • frame_images with a frame_type, for a first or last frame.
  • input_references, for reference media of the types the model lists.
  • generate_audio, where the model offers the choice.
  • callback_url, for a notification when the job finishes.

A conversion step for a deck

For a slide deck, export each slide as an image and pick the few that carry the story. Write the prompt from the slide notes. Pass one or two slides as image references and describe in the prompt what should move. Keep the deck itself out of the request. That gives you a repeatable step that does not depend on a document field the catalog does not list.

What to ask a vendor

When a launch post names a new input type, ask three questions before you design around it. Is the field in the catalog of the platform you call? Is there a size limit that your files exceed? Does the price change when the input is used? The Alibaba post gives a file limit for documents, so the limit exists on Alibaba's cloud, and says nothing about any other route to the same model.

Sources

Related posts

More in Models

All Models posts

Written by Sume