CometAPI’s Wan 3.0 overview describes a document-to-video path that can accept material such as PDFs, presentations, spreadsheets, text or Markdown documents, and webpages, then generate a video intended to reflect the source structure.
That capability creates a new quality problem. A polished explainer can still omit a qualifier, reverse a comparison, display the wrong number, or turn an illustrative graphic into an unsupported conclusion. Visual coherence and factual preservation are separate acceptance gates.
Freeze a fact ledger first
Before generation, extract the claims the video is allowed to make. Each ledger row should contain:
- the exact fact or number;
- its source page, slide, sheet, section, or URL location;
- required qualifiers, unit, date, and comparison baseline;
- whether it must appear in narration, on screen, or both;
- whether it is required, optional, or prohibited from simplification.
Do not ask the model to “summarize accurately” and leave accuracy undefined. For a quarterly report, the acceptance set might require three named metrics, their periods and units, the stated cause of a change, and the source’s own uncertainty language. Material not present in the ledger should not become a new factual claim merely because it makes the story smoother.
Bind the ledger to the exact source revision. If the deck, report, or webpage changes, preserve the earlier ledger and generate a new one rather than reviewing a new video against stale facts. This also makes corrections traceable: the reviewer can distinguish a generation error from a source update that arrived after the first render.
Separate transformation from invention
A document-to-video workflow has several transformation layers: selection, script, scene plan, narration, captions, charts, imagery, and timing. Check each layer against the ledger. A correct script can still become an incorrect chart; an accurate chart can be undermined by narration that implies causation where the source reports correlation.
Visual implication matters too. Showing a rapidly rising line, a confident spokesperson, or a product in a regulated setting can communicate more certainty than the document contains. Reviewers should ask not only “is every sentence true?” but also “does the sequence or image imply an unsupported relationship?”
Run two acceptance passes
The first pass is factual:
- Transcribe narration and on-screen text.
- Match every material statement to a ledger row.
- Verify numbers, units, dates, labels, and qualifiers.
- Record omissions and unsupported additions.
- Reject any scene whose visual implication contradicts the source.
The second pass is temporal and visual. The source warns that a 30-second generation can still suffer identity drift, object deformation, or inconsistent physical relationships. Inspect the full clip, not only attractive keyframes, and keep those defects distinct from factual errors.
The tracked article also describes Wan 3.0 as preview access at its source date, so current availability, input limits, and pricing should be rechecked before a production dependency is made. Preview status does not weaken the editorial standard: if the output is used externally, the facts still need a deterministic review path.
Document-to-video can compress a large source into an accessible format. The compression is precisely why preservation must be tested. Approve the video only when its narration, text, charts, ordering, and visual implications remain inside the source’s factual boundary.