One image, many resolutions, all at once. Blur it, shrink it, repeat — the trail of leftover detail at every step is exactly what lets you rebuild the original perfectly, byte for byte, from a picture the size of a postage stamp.
Blur, halve, repeat
Every level is half the size and a little blurrier than the one below it. What each step throws away is never actually gone — it's saved separately, and that's the whole trick.
Click any level in the staircase to inspect its resolution.
Click a level to see its detail-value histogram — most values cluster near zero, which is exactly why these layers compress so well.
Detail value histogram
Level restored
0/0
Error vs original
—
✅ Perfect reconstruction — pixel-identical to the original
The same real FAST detector runs independently on each pyramid level. Fine detail shows up at the base; broad shape shows up near the top.
Naive blend cross-fades raw pixels at one width for everything. Pyramid blend fades each frequency band separately — the same trick from Lab 24's multi-band blending, but now you can see every band doing its own thing.
Nothing gets lost, it gets filed separately▾
Gaussian pyramid
Blur, then keep every other pixel. Each level is genuinely lower information — this is the "coarse view" half of the pair.
Laplacian pyramid
At every level: what does the ORIGINAL look like minus what you'd get by blowing the next (blurrier) level back up? That difference is the detail this level alone contributes — save all of those, plus the tiny top, and you have lossless storage of the whole image in a different shape.
Why this matters everywhere
Multi-scale feature detection, pyramid optical flow, seamless image blending, and JPEG-style compression all lean on exactly this idea — represent an image as coarse structure plus a stack of detail layers, most of which turn out to be almost entirely zero.
Predict, then verify
You build a 5-level Laplacian pyramid, then reconstruct using only levels 0-3 (skipping the very last, smallest correction layer entirely — pretending it's all zero). How close does the result get to the real original?