Smaller PNGs. Same picture. pngdeer is a lossy PNG compressor with no quality slider: for each image it searches palette sizes from 2 to 256 colors, with and without dithering, scores each result with a perceptual metric, and keeps the smallest file that still looks right. At the same quality, the median image comes out 8–12% smaller than with TinyPNG, the best result of any open-source compressor we've measured.
Try it in your browser → Nothing is uploaded: images are compressed on your device.
brew install heydeer/tap/pngdeer
pngdeer input.png -o output.png- Picks the quality for you: about what TinyPNG would give the image, never above 90 on SSIMULACRA2's scale, which is indistinguishable at 1:1.
- Same bytes everywhere: the command line (macOS and Linux) and the browser (WebAssembly, on every core) run the same Rust code and write identical files.
- Big images too: up to 40 megapixels in the browser, more on the command line.
- Free and open source: MIT.
We ran pngdeer, TinyPNG and eight open-source quantizers on 180 PNGs: icons, UI screenshots, illustrations, photos, cut-outs and 3D renders. Every output was scored against its original with SSIMULACRA2, a perceptual metric where 90 means indistinguishable at 1:1. TinyPNG is the yardstick:
- Points: median SSIMULACRA2 points above (+) or below (−) TinyPNG when a tool's file is as big as TinyPNG's.
- Reaches: images where a tool can match TinyPNG's quality at all.
Each cell gives the held-out set (71 images) / the tuning set (109 images):
| Compressor | Points | Reaches |
|---|---|---|
| pngdeer | +1.7 / +2.0 | 67 / 106 |
| TinyPNG (the yardstick) | 0 | 71 / 109 |
| pngquant | 0.0 / +0.3 | 53 / 87 |
| quantizr | −0.1 / −1.0 | 39 / 53 |
| exoquant | −3.7 / −5.1 | 19 / 24 |
| NeuQuant | −6.5 / −9.8 | 4 / 5 |
| FFmpeg | −7.0 / −8.8 | 16 / 15 |
| quantette | −7.6 / −9.6 | 18 / 21 |
| ImageMagick | −11.3 / −14.5 | 3 / 8 |
| Pillow | −16.8 / −19.5 | 4 / 4 |
At TinyPNG's quality, pngdeer's files are 12% smaller on the held-out set and 8% smaller on the tuning set (median image).
- Held out: pngdeer's parameters were tuned on the 109 images; the other 71 were collected afterwards and never used for tuning.
- Generous to the others: each open-source tool gets its best of ten settings per image, picked after the fact. pngdeer's results are what you get, searched at compression time.
- A second opinion: scored with butteraugli instead, which the search never looks at, pngdeer's files are still 6–9% smaller than TinyPNG's at the same quality.
- TinyPNG ran through its API in October 2026. bench/ reruns everything; the website screenshots in the sets aren't ours to share, so a rerun covers 155 of the 180 images.
- Searches the palette: palette sizes from 2 to 256 colors, each tried a few ways (with and without Floyd–Steinberg dithering, with neighbour smoothing, with a second palette builder), and the smallest file that reaches the target wins.
- Scores what you'd see: SSIMULACRA2, measured in-process. Colors are matched in Oklab, smooth gradients are always dithered, and transparent pixels are judged over light and dark backgrounds.
- Stays fast on big images: large images are searched on a downscaled preview and checked at full size.
The research notes follow every step, from the first comparison with TinyPNG to today's defaults.
pngdeer input.png -o output.png # automatic quality
pngdeer input.png -o output.png --target 85 # a fixed SSIMULACRA2 quality
pngdeer --helpFrom source: cargo build --release, and the binary is target/release/pngdeer.
The web app runs the same code compiled to WebAssembly. To run it locally (web/build.sh lists the toolchain it needs):
sh web/build.sh
python3 web/serve.py # http://localhost:8770MIT. Dependencies keep their own licenses (web/THIRD_PARTY_LICENSES.txt), as do the test and sample images (the SOURCES.md next to them).