How it was made

What was done

Nothing is encoded while you wait. Every file was prepared in advance on one computer, by the programs listed below, and the pages only play them back. Nothing you do here is sent anywhere.

Text. The six compressors were run on the Silesia corpus, a standard collection of real files, joined into a single archive. Each was run on one processor core, and the speeds are wall-clock times for one run.

Audio. Each recording is 12 seconds of 48 kHz, 16-bit stereo. In the viewer every file is decoded by your browser, lined up with the original to the nearest sample, and played in step with it, so that switching changes nothing except the codec. No score is given for audio, because no freely available measure of sound quality can be trusted across codecs. The blind test is the measure.

Image. The photographs are from the Kodak test set. The likeness score is SSIMULACRA 2, in which 90 is visually identical, 70 is high quality, 50 medium and 30 low. Formats are compared by the file size each needs to score 70, read off its curve between the two encodes either side. JPEG XL was measured and is not shown, because few browsers can open it.

Video. Each film is 200 frames at 1280 by 720, 50 frames a second. VMAF was computed against the uncompressed frames. Encoders are compared by the bitrate each needs to reach a VMAF of 85, read off its curve between the two encodes either side. That single point was chosen over an average across the whole curve because VMAF flattens out near the top, where small differences in score stand for large differences in bits. HEVC, VP9 and a second AV1 encoder were measured and are not hosted.

3D. The rabbit was written out as 32-bit positions and 32-bit point numbers, and that plain form is the starting size. Draco was run at its highest compression level on a ladder of grids. The line called grid alone is Draco at level 0, which stores the points and triangles in order with no walk across the surface and no guessing.

Data and DNA. The column format in the Data room and the list of differences in the DNA room were written for this museum, in a few dozen lines each, and every file they produce is decoded again and checked against its source. The differences between sequences were found by edit distance, counting a changed, added or missing letter as one each. The model in the DNA room is not a file. Its size is the length an arithmetic coder would reach using that model's guesses.

What to be careful about

The original in the Video room is a stand-in. Browsers cannot play uncompressed video, so side A is H.264 at a very high quality. On the film with the most noise in it, the stand-in itself scores 95.7 against the true original.

For the two harder films only part of each ladder is hosted, to keep the site a reasonable size. Old town is complete.

Encoding times are single runs on one machine. They are good for comparing one setting with another and for nothing else.

The table of chip makers in the Video room was compiled from their documents and was not measured. The only circuits measured here are those in an Apple M4.

The reference sequence in the DNA room contains one placeholder letter, kept by convention so that positions keep their traditional numbers. It is counted as one of the differences.

The weather table is a reanalysis, a model's reconstruction of past weather from observations, and not the log of a single thermometer. A few readings printed as −0.0 were written as 0.0 before measuring.

A blind test of twelve rounds can show that you hear a difference. It cannot show that there is none.

Programs used

Everything was encoded and timed on one computer (Apple M4, 10 cores) on 2026-11-11.

Material

Every recording, photograph, film, model, table and sequence is public test material, used under the terms given beside it.

Work quoted and not repeated

Some figures in the museum come from published work and were not measured here: Shannon's estimate for English, the two bits a triangle of the walk across a surface, Gorilla's timestamps, Watson's genome, and the language model that beat PNG and FLAC.

Every file and the command that made it

Open a room to see its files. The file names in the commands are the ones used when the museum was built.