What You Can Feed In

Text: the usual way in

Paste it, or drop a file. We read .txt, .docx, .md and .markdown. A .docx is unpacked to plain text in your browser before anything is sent, so formatting, images and comments inside the document are dropped rather than uploaded.

There is a floor: about a paragraph — roughly 330 characters of English prose, or a hundred Chinese characters. Below that there is nothing to break into episodes. If you paste only a link, you will get a specific message saying so — we do not fetch URLs, and a page of ours reading someone else's page is not something you want in the middle of your billing anyway. Paste the text itself.

There is also a ceiling, and it is a hard stop rather than a silent truncation. Long books are handled by reading them in two passes, which raises the ceiling a lot, and the app applies whichever ceiling matches how your text will be read.

We refuse rather than truncate because truncation does not announce itself. The failure we measured on over-long input is a script that comes back looking complete and is not — a fifty-episode book parsed into ten episodes, with no error anywhere. A refusal in the first second is a better outcome than a plausible-looking result you only catch three steps later.

Any language. The interface is separate from the language your production documents come out in, and separate again from the language your characters speak — there is a dedicated guide on that.

Images: reference material and ad photos

Pick whatever your camera or design tool gave you — WebP, HEIC, PNG, JPEG. Your browser decodes it, flattens any transparency onto white and re-encodes it as JPEG before a single byte goes up. One path, no exceptions, including for files that were already fine.

That is why what arrives is always JPEG or PNG: those are the two formats every step downstream accepts, and converting at the door means the server only has to recognise one shape of bytes rather than trust a file extension. Renaming something to .jpg will not get it through.

The only refusals are files your own browser cannot open — HEIC outside Safari, SVG, camera RAW — and you find that out the moment you pick the file, not three steps later.

For an ad, you can attach up to four photos of the same subject — different angles, a detail, the thing in use. Each up to ten megabytes, sixteen in total. They are read once, for free, before anything is charged, and the read tells you what it thinks the product is so you can correct it before spending.

Video: what remaking does and does not do

MP4 or MOV, up to three hundred megabytes, and no longer than the length shown on the upload box. The upload is read: we cut it into shots and describe the people, places and props in it.

What happens next is the part to be precise about. The source video is never handed to the video model. It is used to understand, and what comes out is generated from that understanding — new characters from your setup images, new framing, new lines. The result will not look like the original, and it is not meant to.

Which means the honest use is your own footage, or footage you hold the rights to: a rough cut you want rebuilt, a reference you shot yourself, something you have permission to work from. We do not review uploads before they are processed, so the rights question is yours, and our takedown process applies to generated files and share links the same as anything else.

Reading the video is charged, because it is a real model call over every shot we cut out. The rest of the project is charged step by step afterwards, the same as any drama.

What the three doors have in common

Behind all three is one pipeline: read the material, produce characters, locations and props, break it into episodes, write a storyboard for each, shoot each segment, composite. Ads and remakes simply have something pressing the next button for you.

That is why the editor looks identical whichever door you came in by, and why a guide written about one of them is usually true of the others.

FAQ

Updated October 2, 2026