How our AI works
Most of the AI in these tools runs on your own phone or computer, inside your browser. A few steps use our server and our AI helpers, and each sends only what it needs: the text of your script for videos, your title for thumbnail ideas when you ask, a cut-out picture of your product for the first poster design, and the details about your business for the website maker. Captions, video ads and Instagram cards send nothing.
On your device
When you press a tool's button for the first time, your browser downloads the AI model that tool needs and keeps it. From then on the work happens on your device: the voice is recorded there, the captions are timed there, every frame of the video is drawn there and the MP4 file is put together there.
That is why the site can be free, why there is no sign-up, and why the waiting time depends on your device. The page tells you how long a video will take before you start.
What each tool sends
Video ad maker, poster maker and thumbnail maker: your own photo, once, to cut out its background. We do not keep it.
Script to video and its formats: the text of your script, to plan the pictures and find matching clips, and a topic, if you ask our writer for a script. The video itself is made on your device.
AI captions: nothing. The speech is turned into words on your device.
Thumbnail maker: also your title, but only when you tap "Suggest points from my title".
Poster maker: for the first design only, a picture of your cut-out product and its name, for the AI background. The other designs are made on your device.
Video ad maker: nothing else.
Instagram card maker: nothing. It uses no AI.
Website maker: the details about your business that you type, or your link, so our AI can write the words. Your photos stay on your device.
The script writer
Writing a script from a topic does not run on your device. Writing models small enough for a browser get facts wrong too often, so the writer uses a larger AI: Claude by Anthropic, or open models on Cloudflare's AI service as a backup. Only the topic you type is sent. The script is put into the box so you can read it and change it. AI can still be wrong: check the facts before you post.
In a quiz, questions you type without their answers are sent too, so the writer can add the answers. Nothing else of your script is sent for this.
The stock clips
The moving pictures behind the words are real stock clips from free libraries (Pixabay, Coverr, NASA and Wikimedia Commons). They were filmed by people. An AI (Claude by Anthropic, or Cloudflare's AI service as a backup) reads the whole script and writes the search words for each sentence. Fixed rules then choose the clip or photo: first one that carries the name or the thing the sentence says. No AI picks the picture. The text is not kept. When no clip fits, the video shows a fitting free photo (from libraries such as Unsplash and Pexels) that moves slowly with a sense of depth. When no real photo fits either, it may show a picture made by AI for that sentence (never of holy sites, religious symbols or real people), or the sentence's own key words, a number or a map drawn by our own code. When the AI is not available, our own rules write the search words from the sentence itself. Never a random clip.
The models we use
Voice: Kokoro, an open text-to-speech model (Apache 2.0 licence), with its built-in voices only. Running it: ONNX Runtime and Transformers.js (MIT and Apache 2.0). Putting the MP4 together: your browser's own video encoder, and Mediabunny (MPL 2.0). All of these licences allow free and commercial use.
Two helper parts carry copyleft licences and are shipped as separate parts: eSpeak NG (GPL 3), which turns text into speech sounds for the voice, and the AAC sound encoder from FFmpeg (LGPL), used only where your browser has no AAC encoder of its own.
The German voice is a Piper voice: "Thorsten" (CC0). The thumbnail maker cuts the person out of a photo with MODNet, an open portrait model (Apache 2.0 licence).
Music and moving backgrounds are not AI: they are made by our own code.
Cutting out your own photos: BiRefNet, an open model (MIT licence), run by Cloudflare's image service. As a backup on your device: U-2-Netp and MODNet (both Apache 2.0).
What the AI does not do
It does not copy anyone's voice or face. It does not make AI actors, avatars or generated footage. It does not learn from what you type or send: we do not use the topics, scripts or pictures sent to our server to train anything.
Where it can be wrong
The voice can mispronounce unusual names and numbers. Caption timing is estimated from the speech and can be a fraction of a second early or late. Listen to and watch your video before you post it.
Limits we set on purpose
Scripts with hateful, sexual or dangerous words are refused. Text story videos always carry a small STORY label, so a made-up chat cannot be passed off as a real one.