What this tool reads, what it writes, and what actually makes each one smaller.
PNG is lossless, so the only ways to shrink one are to store the same pixels more cleverly or to store fewer distinct colours. This does both, in your browser.
Re-encode a JPEG at the quality you choose, with the encoder settings that normally sit behind a paywall or aren't offered at all.
WebP usually beats both JPEG and PNG, but not always โ which is exactly why the tool measures instead of assuming.
AVIF is the smallest of the mainstream image formats at a given quality, at the cost of slower encoding and less universal support.
An animated GIF is almost always the wrong format โ converting it to video shrinks it by an order of magnitude.
BMP is uncompressed by design. Practically any other format is smaller, usually by 80โ95%.
Scans and camera TIFFs are enormous. Converting to PNG keeps every pixel; converting to JPEG or WebP keeps the look at a fraction of the size.
Re-encode with H.264, H.265 or AV1, targeting a quality level, a bitrate, or an exact output size.
QuickTime files off a phone or a camera are usually encoded for capture speed, not for size, and re-encoding them wins a lot.
VP9 and AV1 in a WebM container โ the combination that streams smallest on the open web.
Matroska carries almost any codec, so the win depends entirely on what is inside it.
AVI files are usually old and usually encoded with something long superseded. Modern codecs cut them dramatically.
Lower the bitrate, downmix to mono, or move to a modern codec that sounds better at the same size.
WAV is raw PCM. Anything else is smaller โ FLAC losslessly by about half, Opus by ten to twenty times.
Either squeeze the lossless file harder, or accept a lossy codec and save an order of magnitude.
Lower the bitrate, or convert to Opus for better quality at the same size.
Re-encode Vorbis or Opus at a lower bitrate, with the content type set correctly.
STL stores three complete vertices for every triangle and shares nothing, so a typical model repeats each vertex about six times. Removing that costs no shape at all.
OBJ is plain text, which makes it readable and enormous. Converting to a binary mesh format is a large, lossless win.
Scanned meshes are dense far beyond what anyone displays. Decimation and quantisation both apply directly.
The format the web ships 3D in. Reducing the triangle count is what makes a model load quickly on a phone.