Make a video smaller without making it worse.
Pare cuts a video to half its size or less and keeps it looking like the original, or shows how close it came. It encodes across your computer’s cores, in this tab, and the file never leaves your device.
Drop a video anywhere on the pagePick a video from your device
MP4, MOV, WebM, or MKV with H.264, HEVC, VP9, or AV1 video.
Measured on four videos
Visually lossless, at least 50% smaller, in Chrome on a 4-core, 8-thread cloud machine with no GPU. Time runs from clicking Compress to the finished file, and SSIM is averaged over every frame.
| Video | Original | Pare | Time | SSIM |
|---|---|---|---|---|
| Camera footage1080p30, 10 s, H.264 at 62 Mbps | 77.9 MB | 23.9 MB −69% | 11 s | 0.9944 |
| Phone clips1080p50, 20 s, H.264 at 26 Mbps | 65.5 MB | 30.2 MB −54% | 42 s | 0.9505 |
| AnimationBig Buck Bunny, 1080p30, 10 s | 30.7 MB | 14.3 MB −53% | 24 s | 0.9834 |
| Screen recording1080p30, 8 s, scrolling text | 10.8 MB | 3.49 MB −68% | 9 s | 0.9995 |
How it works
Decoded by your browser
WebCodecs decodes the video, on the GPU when there is one, and frames are copied straight into the encoder’s memory.
x264, rebuilt for the web
x264’s speed comes from hand-written x86 and ARM assembly, which a browser can’t run. Pare replaces it with about 1,500 lines of WebAssembly SIMD, bit-exact with x264’s own code and about twice as fast.
Every core busy
The video is split into chunks of equal work, one per encoder, and the encoding starts while you’re still looking at the settings. Short videos get fewer, longer chunks with more threads each, since every chunk begins with a costly keyframe. A chunk that turns out slow hands its last frames to whichever encoder will be free first.
Held to half the size
Short test encodes measure how size falls as quality drops, and each chunk gets its setting from what the finished ones really cost. The file is weighed at the end, and if it isn’t at least 50% smaller, the busiest chunks are encoded again. A video with room to spare gets an x264 preset that does half the work, kept only if VMAF still scores it visually identical.
H.264 or AV1, by measuring
Netflix’s VMAF, compiled to WebAssembly too, scores the test encodes. AV1 is used when it looks better at the target size and your device plays it, or when H.264’s test already shows noise it can’t keep. Otherwise it’s H.264, which is faster and plays everywhere.
Checked frame by frame
Every frame is scored against the source. The result screen puts the weakest frames next to the original so you can judge for yourself.
Read the research notes for the benchmarks behind each choice.