How to convert a video into images only

Sometimes the words are not the point. A screen recording of a settings panel, a webinar where the deck did the talking, a walkthrough with no narration worth keeping: what you want is the pictures, in order, and nothing else.

I learned this by wasting an afternoon. I had a 50 minute screen recording of an internal tool and I ran it through the normal document job. Transcription took most of that time. Then I opened the PDF, read four lines of somebody saying "and then you click here", and knew I would never read another word of it. The only useful thing in the output was the frames folder sitting beside it.

The free way with ffmpeg, and where it stops

If the video is already a file on your disk and you are comfortable in a terminal, ffmpeg will do this today for nothing. Install it, then:

1. Put the video and your terminal in the same folder.
2. Run ffmpeg -i input.mp4 -vf fps=4 -q:v 2 frame_%06d.jpg.
3. Wait. You get four images for every second of video.

That is a real answer and I use it for one-off jobs. Here is where it stops. The names are sequence numbers, so frame_000842.jpg tells you nothing until you divide by four in your head. Every duplicate is kept, so an hour of one static slide becomes 14,400 copies of it. No index, no record of the settings that produced the folder, no way to cover a face. And it needs a file you already have, so a link means downloading it yourself first.

Keyframes, every frame, and the fixed sample

Three different things get called the same thing here, and picking the wrong one is why people end up with the wrong folder.

• Keyframes in the encoder's sense are the frames a compressor stores in full rather than as a difference from the one before. Some tools hand you these because they are cheap to pull out. They are placed to suit compression, not to suit you, and on a still slide there may be one every ten seconds.
• Every single frame means all 30 a second, or 60 on some recordings. It sounds thorough and is mostly waste: an hour is over 100,000 files, and 29 of every 30 are near copies.
• A fixed sample is the useful middle. Four a second is close enough together that nothing meant for a human to read can appear and vanish between two pictures, and an hour lands at about 14,400 before duplicates are removed.

There is a fourth thing, and it is the one you do not want here: the screenshots inside a VideoDoc PDF. Those are chosen. The app scores each picture for content, prefers the slide over the face, collapses repeated slides, and puts what survives next to the words spoken over it. Right output for a document. Wrong output for a raw sequence, because most of the video never makes the cut.

Images only, as its own mode

In VideoDoc this is not a tick box hidden in a panel. It is one of the three things the app can make, chosen at the top: a document for AI, just the video file, or every frame as images. Pick the third and it downloads or opens the video, writes the pictures, and stops.

• You choose 4, 2 or 1 a second, and the app tells you how many images that means per minute and per hour before you start.
• Frames identical to the one before are skipped, which on a slide deck or a screen recording throws away most of the pile.
• Files are named frame_000123_00h01m45.50s.jpg, so the name is the timestamp.
• A frames_index.csv lists every image with its moment, so you can sort and filter and jump back into the video.
• It works on a link and on a file already on your computer, which is where a screen recording usually lives.
• If those pictures are going anywhere near other people, faces can be filled with solid black on the way out.

Nothing is listened to, so the slow half of the job never runs.

What resolution you actually get

The pictures are as big as the video is. Nothing is upscaled, so a 480p file gives 480p images and the small text in them stays unreadable. If you are pasting a link, the quality you pick decides this, and Best fetches the video in full HD. That matters more than it sounds, because the point of these images is usually that something in them can be read, by you or by a model.

They come out as JPEG, not PNG. At the quality level used, screen text looks the same to a human and to an AI, and across 14,400 files the disk cost is not a small difference. If you need lossless frames, this is not your tool, and I would rather say so here than in a refund email.

Why a folder of JPEGs beats the video file for an AI

Dropping an MP4 into a chat window mostly does not work, and where it appears to work the model is usually reading a transcript of the audio rather than watching anything. Images are different. ChatGPT, Claude and Gemini all read images well, and they read a batch of them in order without complaint.

So the folder is the useful shape. Two hundred deduped frames of a screen recording, in order, with the moment in every file name, gets you further than any transcript when the meaning was on screen instead of in the audio. Ask which screen the error appeared on, or ask it to write the steps out from the pictures alone. More on that in feeding a video to ChatGPT or Claude.

Quick questions

Find the screen recording you never rewatched because scrubbing it was too slow, and turn it into a folder of pictures instead.