PNG keeps every pixel. It shrinks files by predicting the next one.
PNG never throws detail away. Instead, each row of pixels is rewritten as the small difference from what was expected, then zipped. Flat colours and sharp edges predict well, so screenshots and logos shrink; photos are full of surprises, so they stay huge. Pick an image and see the prediction at work.
Design tools offer several kinds of PNG. They differ in how many colours each pixel can hold and whether it can be see-through. Picking the smallest one that fits is the easiest size win.
Every PNG begins with the same 8 bytes, then a series of chunks, each named with four letters. Only three are required. Screenshots and exports often carry extras: colour profiles, text notes, software names. Removing them is how "PNG optimisers" save space without touching a pixel.
This page walks every chunk, recomputes each CRC-32, inflates the image data with your browser’s own DEFLATE, checks the Adler-32 and undoes the row filters itself. So it can show which filter each row used, what the file carries besides pixels, and whether the pixel format is bigger than the content needs. Everything happens on your device. Both samples are real PNGs built in your browser; the second is broken on purpose.
Identification, limits and capabilities, taken from the W3C specification.
Every PNG opens with the same 8 bytes. Each one is there for a reason: together they detect files mangled by text-mode FTP, 7-bit email and line-ending conversion, which were everyday problems in 1996. Click a byte.
89 50 4E 47 0D 0A 1A 0A signature00 00 00 0D IHDR length: 13 bytes49 48 44 52 "IHDR"00 00 04 00 00 00 03 00 width 1024 · height 76808 06 8-bit · RGBA00 00 00 deflate · adaptive · no interlaceB7 5C 42 11 CRC-32 of the chunk
When Unisys began enforcing its patent on GIF’s compression, a group on Usenet designed a free replacement. Thirty years later, the third edition added HDR and animation.
Each converter page explains what changes on the way.