JPEG Compression: Full-Image DCT Tiling & File-Size Estimate (2D)

Source Image

Compression

Quality factor 50

Live Stats

Raw size (grayscale)

–

Est. JPEG size

–

Compression ratio

–

PSNR (whole image)

–

Avg. surviving AC coeffs / 8×8 block (of 63)

–

How it works

The image is split into non-overlapping 8×8 pixel blocks. Each block runs through the exact pipeline JPEG uses on its luminance channel:

forward:  X = DCT2D(pixels − 128)
quantize: Q[i] = round(X[i] / step[i](quality))
decode:   pixels' = IDCT2D(Q · step) + 128

The step-size table is the standard JPEG luminance quantization matrix, scaled by the quality factor — low quality means large steps, so more high-frequency coefficients round to exactly zero. That is the actual mechanism of lossy compression: not blurring, but discarding coefficients.

The estimated JPEG size is computed per block from how many quantized coefficients survive and their magnitude (rough bits-per-coefficient model, plus an end-of-block marker and a small header) — it is an approximation of real Huffman-coded JPEG, not a byte-exact encoder, but it tracks the same trend: fewer / smaller surviving coefficients → smaller file.

Hover or drag over either image to open a 16×16-pixel loupe zoomed 10×, so you can see individual 8×8 block boundaries and ringing/blocking artefacts directly on the pixels that produce them.
Checkerboard and Noise patterns spread energy across all 64 coefficients almost equally — DCT cannot concentrate energy in a few low frequencies the way it does for smooth gradients, so they compress far worse at the same quality.
Both images are compared as grayscale (luminance-only) so the size/PSNR numbers reflect the transform itself, not chroma subsampling.

Original (grayscale)

Compressed (reconstructed)