How to turn a book into a video without rewriting a word

Paste a chapter into almost any AI video tool and watch what comes back. The story is recognisable. The pacing is tighter. Somewhere in the second scene a sentence you spent an afternoon on has become a shorter, flatter sentence that means roughly the same thing.

Nobody chose this. It falls out of how these tools are built. The standard design asks a language model to read your text and return a structured scene list, because that is the format the renderer needs. A model that receives prose and emits prose rewrites the prose. Not out of malice, and not because the prompt failed to say otherwise. Generating text is the entire operation being performed.

For most uses this is fine or even wanted. You hand over a topic and you want the model to write something. For an author it is the one outcome that makes the whole thing worthless. A picture book is not a summary of its plot, and a paragraph that took a week is not interchangeable with a paragraph that means the same thing.

Why telling the model not to rewrite does not work

The obvious fix is an instruction: return the text unchanged, copy it exactly, do not alter a single word. It works most of the time, which is the problem. Most of the time is a property you audit after generation rather than one you can promise, and the failures are quiet. A dropped clause in scene nineteen of sixty does not announce itself.

Once you accept that, the fix stops being a better prompt and becomes a different architecture. If no model ever sees the narration, there is no mechanism by which a word can change, and nothing downstream needs checking.

Three jobs have to be separated that most pipelines do in one call.

• Cutting the text into scenes. Pure code, no model. The text is split on sentence boundaries and sentences are packed into runs sized for the screen. Concatenating the runs reproduces the book exactly, because nothing else ever happened to it.
• Deciding what each scene looks like. This is where a model belongs. It receives one finished run of the author's text and answers a single question: what should this frame show. It returns an image description and never returns narration.
• Reading the book once, for context. One pass over the story establishes the title, the audience, the genre, the tone, the recurring characters and a single sentence describing how every frame should look. This is metadata about the book. It is not the book.

In AutoVidGen the segmentation targets scenes of roughly twelve seconds of narration, with a floor of five and a ceiling of fourteen. Those numbers are not arbitrary and they are not about books. They are what the renderer's motion is tuned for: a three-second still reads as a flicker, and a thirty-second one reads as a slideshow that stopped.

Sentence splitting is where this quietly breaks

Splitting on periods is the version everyone writes first, and it fails immediately on exactly the books this is for. Children's books are dialogue-heavy, and dialogue is full of the cases a naive split gets wrong.

This is the question authors ask first and most tools answer worst, so it is worth being specific.

When you upload a PDF, the file does not go to our server. The text is pulled out of it inside your own browser, in a background worker, and the only thing that crosses the network is the resulting string. There is no copy of your manuscript on a disk somewhere, because the manuscript was never sent.

That decision was made for security rather than for marketing. A document parser accepting arbitrary uploads is one of the more attractive things you can put on a server, and the safest version of that component is the one that does not exist. The side effect happens to be the answer an unpublished author wants.

• What you can upload. PDF, plain text and Markdown files, up to 25 MB. You can also paste the story straight into the box, which is the fastest route for anything short.
• What is not supported yet. EPUB and Word documents. The app says so when you pick one and tells you to open it and paste the text instead, rather than failing with something vague.
• Scanned books. A PDF that is photographs of pages contains no text to extract. This is detected on upload and named immediately, instead of surfacing three screens later as a confusing complaint that your story is too short.

A book generates a lot of scenes. Two thousand words is somewhere around seventy of them, and every one needs an image. If the hero is generated fresh each time, you get seventy different children and a video nobody can follow.

The fix is that the characters are read out of the book once, before any image exists. The pass that establishes the title and tone also extracts the recurring cast, up to six leads, each with a description taken from the author's own text rather than invented. Every scene featuring that character is generated against the same fixed description.

This is the part where the source material genuinely helps. Most AI video has to invent a character and then fight to keep it stable. A book already told you what the fox looks like, in the author's words, and those words are the same on scene sixty as on scene one.

The style anchor

Alongside the cast, that same pass writes one sentence describing how every frame of this particular film should look: the medium, the palette, the light. That sentence is prepended to every image prompt in the video, including the cover.

It is the same principle as fixing a style block across an entire series, applied to a single book. Consistency of look is not a per-scene decision, and anything decided per scene will drift.

The prompting logic underneath this is the same one that makes a series look like a series: why a fixed style block beats a per-scene one

The honest section, because the limits here are real and finding them after uploading is a bad experience.

Two thousand two hundred and fifty words is a picture book, a short story, a chapter, an essay. It is not a novel. A forty thousand word manuscript is refused, and the refusal tells you how many parts it would take rather than just saying no, because a novel is not a mistake somebody made. It is eighteen episodes.

Splitting long books into a series automatically is not built yet. Every video already records which slice of which book it came from, so the groundwork is there, but today a long book means uploading it in pieces yourself.

It is built for people who already have the words. That sounds obvious and it rules out most of the audience for AI video tools, who arrive with a topic rather than a text.

• Children's authors. The clearest fit. Picture books sit inside the word ceiling comfortably, the cast is small and well described, and a read-along video is a format the audience already watches.
• Self-published and unpublished writers. A short story or an opening chapter as a video is a promotional asset that would otherwise cost a narrator, an illustrator and an editor.
• Anyone sitting on finished writing. Essays, folk tales, translations, a grandparent's memoir. Anything already written in sentences somebody cared about.
• Not for topic-first video. If you want a video about a subject and have not written it yet, this is the wrong pipeline. The other ones write the script for you, which is the right behaviour when there is no manuscript to protect.

The pitch is narrow on purpose. Filming a book word for word is worth building only if the words are the point, and if the words are not the point you are better served by a pipeline that writes them.

Everything after the segmentation step is the same pipeline the rest of the product uses: how a finished video gets made, scene by scene