Can AI turn a spreadsheet or slide deck into a video?

Yes. Alibaba's Wan 3.0 accepts documents — spreadsheets, slide decks, PDFs, markdown files, and live web pages — as direct input and builds a video from the source content. The model reads the file, extracts what the numbers mean, and decides what earns the screen, all without a text prompt or reference image describing a shot.

How does Wan 3.0's document-to-video feature work?

Instead of rendering a described shot, Wan 3.0 reads the document first and then invents the intent and the frame together. It decides the color of the room, the focal length, the framing, and the pace of the cut — none of which exist anywhere in the file. Those creative decisions come from the statistical center of everything the model has seen, not from the user.

Do AI video generators still need text prompts?

Not necessarily. Prompt-based generation still works — a filmmaker can write a specific prompt and get a specific frame back, and Wan 3.0 keeps that path open. Its document input adds a second path where no description ever existed, which means the result is always the model's version of the thing rather than anyone's.

Nobody described anything -- CinePrompt Field Notes

Alibaba released Wan 3.0 this week. The name will not stop anyone scrolling, and neither will most of the feature list. Thirty seconds in a single continuous shot. Characters that stay consistent. Camera moves that respect where a director places the lens. Every video model claims these now, and every one of them delivers them imperfectly. Then there is the input that is genuinely new.

The model accepts documents. Spreadsheets, slide decks, PDFs, markdown files, live web pages. You hand it a deck and it reads the source content and builds a video from it. Not a video the deck describes. A video the deck becomes.

For three years the input to these tools has been a description of a shot. A text prompt is a description, however vague. A reference image is a description written in pixels. An avatar is a description of a person. Even an agent that writes the prompt for you still carries the trace of someone who wanted something. The document is the first input that contains no description at all.

A spreadsheet does not describe a shot. It records numbers. A revenue forecast does not say where the light falls, or how low the camera sits, or whether the frame should hold empty for a beat. None of those decisions are in the file. The model makes all of them, silently, from the statistical center of everything it has seen.

The tool stopped asking you to describe anything. That is the change, and it is larger than the feature list suggests. The pipeline this entire discipline runs on goes from intent to image. Someone has a shot in their head. They find the words or the reference to carry it. The model renders. A document reverses the direction. There is no shot in the head. There is a record. The model invents the intent and the frame together.

For a hundred years, film required a person to want a specific frame. Every other input these tools accept carries that want, however clumsily it travels. The document is the first that does not, and that is exactly its selling point. Alibaba did not build this for the filmmaker who knows what the light should look like. It built it for the person who has forty slides and a deadline and no particular feeling about how the camera should move.

That person exists in staggering numbers. There are more people holding quarterly reports than holding shot lists. The market for turning a deck into a thirty-second film will dwarf the market for finding the right aperture, and the model maker knows it. The feature is not a dumbing down of the tool. It is a discovery of who most of the users were all along.

The comprehension part is real, and it deserves its sentence. To turn a deck into a film, the model has to read it first, extract what the numbers mean, decide what earns the screen. That is not generation. It is reading, and a model that reads before it renders is a serious thing. The point is not that the capability is hollow. The point is who it hands the artistic decisions to.

Consider what the model does when it turns that spreadsheet into a film. It decides the color of the room. It decides the focal length, the framing, the pace of the cut, whether the camera pushes in or sits still. Every one of those decisions is a creative act, and every one of them is now performed by no one on purpose. The numbers carry no opinion about any of it. What returns is not a filmmaker's vision realized. It is the model's opinion of what a spreadsheet should look like when it moves, wearing the numbers like a costume.

The same model family has been in the field for a while, and directors have been describing precise shots into it the whole time. A filmmaker can still write a specific prompt and get a specific frame back. The document input does not remove that. It adds a second room next to it, one where no description ever existed, and which will therefore always produce the model's version of the thing rather than anyone's.

There is something faintly comic about watching this arrive in the same season the true cost of a good prompt finally got named out loud. The industry spent two years insisting the sentence was the whole product. Now it ships a feature that works because there is no sentence.

The intent was never in the file.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never made a film from a spreadsheet and suspects the day it becomes necessary will find him elsewhere.