Input 0 BH 0 bit/symL̄ 0 bit/symRatio 0×Now —FPS 0

Left: input bits (digits = the real bits of your data). Centre: the compressor turns them into a core of cubes. Right: three output rows — blue = Huffman, violet = LZ77, amber = LZW; lit cells = compressed size relative to the input (192 lit of 192 = no saving).

Crystal Compressor: Entropy, Huffman, LZ and Loss

Watch real bits flow through a compressor and see where the redundancy goes. Everything runs in your browser; your data never leaves the page.

1. Choose data

or a small file (≤ 16 KB, stays local)

2. Choose codec and run

Results

Symbol frequencies and Huffman code

Lossy: 64×64 picture through 8×8 DCT blocks

Left: original. Right: after quantisation. Move the slider to trade size for quality.

Rate–distortion limit (Gaussian source)

Missions

Independent numerical checks

Also run outside the browser (node src/check.js): round-trips on every source and a comparison with zlib-9 (ours is within 1.2× on English text, about 2× on the repetitive text because zlib also entropy-codes its tokens).

Educational model. Sizes are bit counts of simplified formats (a small header is included and described below); real file formats such as DEFLATE or JPEG add more stages.

Controls

Drag / one finger: orbit; wheel / pinch: zoom; buttons Hero / Side / Top / Inside jump to preset cameras.
Keys: Space pause, S step, R reverse (decompress), 1-4 cameras.
The speed slider sets symbols per second; the pulse in the core grows with the compression ratio.
A small cube core = a big saving; a core as big as the input = nothing was saved.

The science

Entropy. For symbol probabilities p_i, H = −Σ p_i log₂ p_i bits/symbol is the average surprise. No lossless code can have a mean length below H (source-coding theorem).

Huffman. Repeatedly merge the two rarest symbols; code lengths l_i satisfy Kraft's inequality Σ 2^−l_i ≤ 1 and H ≤ L̄ < H + 1. When all probabilities are powers of 1/2 (the "dyadic" source) L̄ = H exactly. Header cost used here: 16 bits + 13 bits per distinct symbol (a model value).

LZ77. A window of W bytes and a look-ahead of B bytes; each token is (offset, length, next byte) and costs ⌈log₂ W⌉ + ⌈log₂ B⌉ + 8 bits. LZW grows a dictionary (max 4096 entries) and emits codes of max(9, ⌈log₂ |D|⌉) bits. RLE stores (byte, run) pairs at 16 bits per run.

Lossy. Each 8×8 block is transformed with the 2-D DCT, coefficients are divided by s·Q₀ (standard luminance table scaled by the quality factor s) and rounded; size is the Huffman-coded quantised stream. PSNR = 10 log₁₀(255²/MSE) dB.

Rate–distortion. For a Gaussian source with variance σ² the best possible rate at distortion D is R(D) = ½ log₂(σ²/D). A uniform quantiser with entropy coding sits about ½ log₂(2πe/12) ≈ 0.254 bit above it; the page measures this on 200,000 samples.

Why random data does not compress. Its entropy already equals the symbol width (8 bit/byte), so there is no redundancy to remove; the headers make the result slightly larger. Kolmogorov complexity states the same idea without probabilities but cannot be computed.

Honest simplifications. No full DEFLATE or JPEG; no arithmetic coder (ideal arithmetic coding would approach H + ~2 bits total); greedy LZ77 parsing; image limited to 64×64 greyscale; input limited to 16 KB; the codec step counts are not timings. Sources: information-theory textbooks (Shannon 1948, Cover & Thomas), the DEFLATE and JPEG standards (by type). See also the Huffman and LZ77 pages of this site.

Models built with Blender (compressor frame, glass, blocks, floor) and rendered with Three.js. Not affiliated with any codec or company.