Input token
Also known as: prompt token, input unit
The unit a model measures its input in, which for video analysis is dominated by how many frames and how large they are.
Updated
What it means
An input token is a small piece of whatever goes into a model: a fragment of a word, or for images a block of pixels. The reason anybody editing video cares is that tokens are the meter.
Everything a model is shown is counted in them, and a single frame is worth hundreds, so frames dominate the total on any video job while the wording of the question barely registers.
Thinking in tokens changes how analysis is designed. The question stops being how long the footage is and becomes how much of it needs to be seen in order to judge it, which is what Frame sampling answers.
How it works in practice
The main lever is sampling rate. A frame every few seconds, chosen for being sharp and representative, describes a clip at a tiny fraction of the full count without losing much that matters.
Resolution is the second lever. A frame sent at a quarter of its size uses far fewer tokens and still supports most judgements about what is happening, so analysis almost never looks at full 4K video frames.
Choosing which frames to send is the third. Picking deliberately, as Automatic keyframe selection does, buys more information per token than raising the rate and sending more blurred duplicates.
What to watch out for
Filling the input is not the same as informing the model. Hundreds of near identical frames from a static shot add count without adding anything to the answer, and they crowd out frames that would have helped.
Token totals also interact with the Context window. A sample that fits comfortably for a short clip overflows on a long one, so the sampling rate has to scale with duration rather than stay fixed.
Audio is easy to forget in the arithmetic. A soundtrack is cheap per second compared with frames, which means it is usually worth sending in full rather than sampling it the same way.
How RawClip handles it
What reaches the model is a sampled, downscaled set of frames rather than whole files, which keeps the analysis of a multi hour upload within reach while still giving it a look at every part of the footage.