Assay · macOS · Built soloNot released yet
It compresses video and picks the settings itself.
You get one knob, size against quality.
Where it came from
This one came out of cleaning a disk. I had 144 screen recordings taking 34.6 gigabytes, and I went looking for why ninety seconds of a call weighs 654 megabytes. Turns out a Mac records the screen at 2880 by 1800, 60 frames a second, 57 megabits. For two talking heads that is generous.
Existing compressors ask you things you do not know. Resolution, bitrate, codec, profile. Assay comes at it from the other end. It measures the file first, how much motion, fine detail, smoothness and sound are in there, and derives the settings from that. What is left for you is one slider with five steps, from nearly untouched to as small as it goes.
The name comes from assaying, where you determine what a metal is made of before you judge what it is worth. Same order of operations.
How it works
The size estimate comes from a sample of the file itself. Three three-second chunks of the actual file get encoded with the actual settings, all five steps in a single pass. After that it is multiplication.
Two axes of quality instead of one. Measuring a still frame tells you nothing about smoothness, so the second axis works the same way as the first. Frames get thinned out, duplicated back in, and compared against the original run.
The original is never deleted. Encoding goes to a temporary file, and the finished result is verified before the program reports anything as done.
There are exactly two codecs. libx264 and aac. No h265, no AV1, because the result has to open everywhere without a conversation.
Measurements
- Calibrated on 142 of my own recordings on 4 August 2026. 33.29 gigabytes came out as 1.26, a factor of 26.4.
- The spread is enormous. Lowest 2.2 times, median 23.3, highest 61.
- Between the lightest and the heaviest kind of content there is a fivefold difference in how many bits it wants. Which is why one preset for everything cannot work.
- The pre-encode size estimate is off by 10.1 percent at the median, against 44.3 percent for the naive method.
- A 4.5 gigabyte, 11 minute file is analysed in 5 seconds. The full five-step ladder on a 624 megabyte file takes 22.
What has not been checked
I have never once listened back to verify the speech. I looked at exactly one file with my own eyes, the worst in the set, and I never looked at smoothness at all.
The thresholds are calibrated on Mac screen recordings between 2466 and 2880 pixels wide. There was no camera footage and no drawn animation in the set, so behaviour on those is simply unknown.
When you can try it
Right now it lives on my own machine. There are no public releases yet, and putting a button here would be dishonest. When there is a build worth handing over, the link goes in this spot.