A simpler wrapper around qoi (https://github.com/phoboslab/qoi)

These details have not been verified by PyPI

Project links

Project description

QOI

A simple Python wrapper around qoi, the "Quite OK Image" image format. It's

Lossless with comparable compression to PNG, but fast! It encodes 10x faster and decodes around 5x faster than PNG in OpenCV or PIL.
You can make it lossy with a simple trick, and then you can get similar compression to JPEG but a whole lot faster. (These number vary a lot depending on how "lossy" you make JPEG or QOI). That's cool.
Multi-threaded - no GIL hold-ups here.

Install

pip install qoi

Example

import numpy as np
import qoi

# Get your image as a numpy array (OpenCV, Pillow, etc. but here we just create a bunch of noise)
rgb = np.random.randint(low=0, high=255, size=(224, 244, 3)).astype(np.uint8)

# Write it:
_ = qoi.write("/tmp/img.qoi", rgb)

# Read it and check it matches (it should, as we're lossless)
rgb_read = qoi.read("/tmp/img.qoi")
assert np.array_equal(rgb, rgb_read)

# Likewise for encode/decode to/from bytes:
bites = qoi.encode(rgb)
rgb_decoded = qoi.decode(bites)
assert np.array_equal(rgb, rgb_decoded)

# Benchmarking
from qoi.benchmark import benchmark
benchmark()  # Check out the arguments if you're interested

If you want to really max out your CPU:

from concurrent.futures import ThreadPoolExecutor, wait
import numpy as np
import qoi

RGB = np.random.randint(low=0, high=255, size=(224, 244, 3)).astype(np.uint8)

def worker():
    bites = bytearray(qoi.encode(RGB))
    img_decoded = qoi.decode(bites)

print("Go watch your CPU utilization ...")
with ThreadPoolExecutor(8) as pool:
    futures = [pool.submit(worker) for _ in range(10000)]
    wait(futures)

(Single-threaded) Benchmarks

If we consider lossless, then we're generally comparing with PNG. Yup, there are others, but they're not as common. Benchmarks:

Test image	Method	Format	Input (kb)	Encode (ms)	Encode (kb)	Decode (ms)	SSIM
all black ('best' case)	PIL	png	6075.0	37.75	6.0	16.04	1.00
all black ('best' case)	opencv	png	6075.0	23.82	7.7	17.93	1.00
all black ('best' case)	qoi	qoi	6075.0	4.13	32.7	2.67	1.00
koi photo	PIL	png	6075.0	849.07	2821.5	85.46	1.00
koi photo	opencv	png	6075.0	95.24	3121.5	44.34	1.00
koi photo	qoi	qoi	6075.0	28.37	3489.0	17.19	1.00
random noise (worst case)	PIL	png	6075.0	300.37	6084.5	46.30	1.00
random noise (worst case)	opencv	png	6075.0	63.72	6086.9	14.01	1.00
random noise (worst case)	qoi	qoi	6075.0	16.16	8096.1	7.67	1.00

So qoi isn't far off PNG in terms of compression, but 4x-20x faster to encode and 1.5x-6x faster to decode.

NB:

There's additional overhead here with PIL images being converted back to an array as the return type, to be consistent. In some sense, this isn't fair, as PIL will be faster if you're dealing with PIL images. On the other hand, if your common use case involves arrays (e.g. for computer vision) then it's reasonable.

Produced with python src/qoi/benchmark.py --implementations=qoi,opencv,pil --formats=png,qoi on an i7-9750H. Not going to the point of optimised OpenCV/PIL (e.g. SIMD, or pillow-simd) as the results are clear enough for this 'normal' scenario. If you want to dig further, go for it! You can easily run these tests yourself.

If we consider lossy compression, again, JPEG is usually what we're comparing with. Normally, it'd be unfair to compare QOI with JPEG as QOI is lossless, however we can do a slight trick to make QOI lossy - downscale the image, then encode it, then upsample it by the same amount after decoding. You can see we've implemented that below with a downscaling to 40% and JPEG quality of 80 (which results in them having the same visual compression i.e. SSIM). So, results (only on koi photo as the rest are less meaningful/fair for lossy):

Test image	Method	Format	Input (kb)	Encode (ms)	Encode (kb)	Decode (ms)	SSIM
koi photo	PIL	jpg @ 80	6075.0	47.67	275.2	24.01	0.94
koi photo	opencv	jpg @ 80	6075.0	24.03	275.3	19.58	0.94
koi photo	qoi	qoi	6075.0	23.17	3489.0	12.94	1.00
koi photo	qoi-lossy-0.40x0.40	qoi	6075.0	4.38	667.5	2.96	0.94

Here we see that lossless qoi is losing out considerably in compression, as expected for lossy vs lossless. Also, qoi is only 1x-2x faster of encoding, and 1.5x-2x faster for decoding. However, it's important to note that this varies a lot depending on the jpeg quality specified - here it's 80 but the default for OpenCV is actually 95 which is 3x worse compression and a bit slower.

However, that's still lossy vs lossless! If you look at qoi-lossy-0.40x0.40 where we downscale as above, you can see that it can perform really well. The compression ratio is now only 3x that of JPEG (and 5x better than lossless QOI, and also the same as the default OpenCV JPEG encoding at a quality of 95), but it's so fast - 5x-10x faster encoding, and 7x-8x faster decoding.

Anyway, there are definitely use cases where qoi may still make sense over JPEG. Even lossless QOI can be worth it if size isn't an issue, as it's a bit faster. But if you use the "lossy" QOI, you're getting "comparable" (depending on JPEG quality) compression but much faster.

NB:

See above re additional PIL overhead.

Produced with python src/qoi/benchmark.py --images=koi --implementations=qoi,qoi-lossy,opencv,pil --formats=jpg,qoi --qoi-lossy-scale=0.4 --jpeg-quality=0.8 on an i7-9750H. Not going to the point of optimised OpenCV/PIL (e.g. SIMD, or pillow-simd, libjpeg-turbo, different JPEG qualities, etc.) as the results are clear enough for this 'normal' scenario. If you want to dig further, go for it! You can easily run these tests yourself.

Developing

git clone --recursive https://github.com/kodonnell/qoi/
USE_CYTHON=1 pip install -e .[dev]
pytest .

We use cibuildwheel to build all the wheels, which runs in a Github action. If you want to check this succeeds locally, you can try (untested):

cibuildwheel --platform linux .

Finally, when you're happy, submit a PR.

Publishing

When you're on main on your local, git tag vX.X.X then git push origin vX.X.X. This pushes the tag which triggers the full GitHub Action and:

Builds source distribution and wheels (for various platforms)
Pushes to PyPI
Creates a new release with the appropriate artifacts attached.

TODO

Make benchmark.py a CLI in setup.py
Create a qoi CLI
Benchmarks - add real images, and also compare performance with QOI to see overhead of python wrapper.
setuptools_scm_git_archive?
Code completion?

Discussion

Wrap or rewrite?

For now, this is just a simple wrapper. We'll leave the original project to do all the hard work on performance etc., and also maintaining (or not) compatibility or adding new features etc. We make no claims to do any more than that - we're basically just porting that C functionality to (C)Python.

On the name

For now, let's rock with qoi because

We're already in python, and the py in pyqoi seems redundant. For what it's worth, 3 < 5.
pyqoi seems like a good name for a python-only version of QOI (useful for pypy etc.), which this isn't.
qoi is generally new so let's not overthink it for now. We can always rename later if needed.

What's up with `./src`?

See here and here. I didn't read all of it, but yeh, import qoi is annoying when there's also a folder called qoi.

`USE_CYTHON=1`?

See here. Fair point.

Project details

These details have not been verified by PyPI

Project links

Release history Release notifications | RSS feed

0.7.0

Jun 23, 2024

0.6.0

Apr 20, 2024

0.5.0

Dec 13, 2023

This version

0.4.0

Dec 11, 2023

0.3.1

Mar 31, 2023

0.3.0

Mar 17, 2023

0.2.0

Jul 21, 2022

0.1.2

Jun 29, 2022

0.1.1

Jun 5, 2022

0.0.9

Dec 30, 2021

0.0.8

Dec 29, 2021

0.0.6

Dec 21, 2021

0.0.5

Dec 20, 2021

0.0.4

Dec 6, 2021

0.0.3

Dec 6, 2021

0.0.2

Dec 6, 2021

0.0.1

Dec 5, 2021

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

qoi-0.4.0.tar.gz (3.3 MB view details)

Uploaded Dec 11, 2023 Source

Built Distributions

qoi-0.4.0-cp312-cp312-win_amd64.whl (90.6 kB view details)

Uploaded Dec 11, 2023 CPython 3.12 Windows x86-64

qoi-0.4.0-cp312-cp312-win32.whl (77.7 kB view details)

Uploaded Dec 11, 2023 CPython 3.12 Windows x86

qoi-0.4.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (519.1 kB view details)

Uploaded Dec 11, 2023 CPython 3.12 manylinux: glibc 2.17+ x86-64

qoi-0.4.0-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl (492.0 kB view details)

Uploaded Dec 11, 2023 CPython 3.12 manylinux: glibc 2.17+ i686 manylinux: glibc 2.5+ i686

qoi-0.4.0-cp312-cp312-macosx_10_9_x86_64.whl (102.0 kB view details)

Uploaded Dec 11, 2023 CPython 3.12 macOS 10.9+ x86-64

qoi-0.4.0-cp311-cp311-win_amd64.whl (89.7 kB view details)

Uploaded Dec 11, 2023 CPython 3.11 Windows x86-64

qoi-0.4.0-cp311-cp311-win32.whl (76.9 kB view details)

Uploaded Dec 11, 2023 CPython 3.11 Windows x86

qoi-0.4.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (526.3 kB view details)

Uploaded Dec 11, 2023 CPython 3.11 manylinux: glibc 2.17+ x86-64

qoi-0.4.0-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl (503.4 kB view details)

Uploaded Dec 11, 2023 CPython 3.11 manylinux: glibc 2.17+ i686 manylinux: glibc 2.5+ i686

qoi-0.4.0-cp311-cp311-macosx_10_9_x86_64.whl (100.6 kB view details)

Uploaded Dec 11, 2023 CPython 3.11 macOS 10.9+ x86-64

qoi-0.4.0-cp310-cp310-win_amd64.whl (89.7 kB view details)

Uploaded Dec 11, 2023 CPython 3.10 Windows x86-64

qoi-0.4.0-cp310-cp310-win32.whl (77.1 kB view details)

Uploaded Dec 11, 2023 CPython 3.10 Windows x86

qoi-0.4.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (488.1 kB view details)

Uploaded Dec 11, 2023 CPython 3.10 manylinux: glibc 2.17+ x86-64

qoi-0.4.0-cp310-cp310-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl (468.3 kB view details)

Uploaded Dec 11, 2023 CPython 3.10 manylinux: glibc 2.17+ i686 manylinux: glibc 2.5+ i686

qoi-0.4.0-cp310-cp310-macosx_10_9_x86_64.whl (100.7 kB view details)

Uploaded Dec 11, 2023 CPython 3.10 macOS 10.9+ x86-64

qoi-0.4.0-cp39-cp39-win_amd64.whl (90.2 kB view details)

Uploaded Dec 11, 2023 CPython 3.9 Windows x86-64

qoi-0.4.0-cp39-cp39-win32.whl (77.8 kB view details)

Uploaded Dec 11, 2023 CPython 3.9 Windows x86

qoi-0.4.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (490.6 kB view details)

Uploaded Dec 11, 2023 CPython 3.9 manylinux: glibc 2.17+ x86-64

qoi-0.4.0-cp39-cp39-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl (470.9 kB view details)

Uploaded Dec 11, 2023 CPython 3.9 manylinux: glibc 2.17+ i686 manylinux: glibc 2.5+ i686

qoi-0.4.0-cp39-cp39-macosx_10_9_x86_64.whl (101.3 kB view details)

Uploaded Dec 11, 2023 CPython 3.9 macOS 10.9+ x86-64

qoi-0.4.0-cp38-cp38-win_amd64.whl (90.3 kB view details)

Uploaded Dec 11, 2023 CPython 3.8 Windows x86-64

qoi-0.4.0-cp38-cp38-win32.whl (77.8 kB view details)

Uploaded Dec 11, 2023 CPython 3.8 Windows x86

qoi-0.4.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (498.9 kB view details)

Uploaded Dec 11, 2023 CPython 3.8 manylinux: glibc 2.17+ x86-64

qoi-0.4.0-cp38-cp38-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl (477.5 kB view details)

Uploaded Dec 11, 2023 CPython 3.8 manylinux: glibc 2.17+ i686 manylinux: glibc 2.5+ i686

qoi-0.4.0-cp38-cp38-macosx_10_9_x86_64.whl (100.5 kB view details)

Uploaded Dec 11, 2023 CPython 3.8 macOS 10.9+ x86-64

File details

Details for the file qoi-0.4.0.tar.gz.

File metadata

Download URL: qoi-0.4.0.tar.gz
Upload date: Dec 11, 2023
Size: 3.3 MB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: twine/4.0.2 CPython/3.9.18

File hashes

Hashes for qoi-0.4.0.tar.gz
Algorithm	Hash digest
SHA256	`f0715587e57944a4c0b9c3c7c54b45f7e4da29ef666ffcf75d26e286167f2266`
MD5	`332a8a7f4bc63d25227c5c19c640d12e`
BLAKE2b-256	`4f609818b97b0517e197c9149cc35eb0767d188b8c9e811c4ac6684a7e9a6288`

Algorithm	Hash digest
SHA256	`05dc29e5e0035ed0075a771eca8f9ee58fe417feae652b6ee2326ce98d7cb9a9`
MD5	`30f0f17b1dd65b99742767762cb49d55`
BLAKE2b-256	`e05ef98e244e244573f2a291cbcb310a0890c19e364227b4fc0c96759aa8426f`

Algorithm	Hash digest
SHA256	`35dd657168dec69ccda4b0257fee078d5b3d305d759367ac8bff361f7d76e59b`
MD5	`ffda41ecbafe3621a431ee2ef6bd79e0`
BLAKE2b-256	`abde6041f8c2d166d3c1eaf900ae000406630e1756de6cb93e3d38076d435db6`

Algorithm	Hash digest
SHA256	`8c85517211fcf7e9f30740fe2395d4c3db51ced84c85479b2aeee4204bb057b2`
MD5	`ef8e3d401b9f0e429514d6b49ffcdfa3`
BLAKE2b-256	`c8530eafad01f8b2a474ea30718dc84f21b4a6639cdba683ac06129657fc241d`

Algorithm	Hash digest
SHA256	`712574054e76fd5dc49436392f4682c76b3bbcda242d12e1dda2f2fe4eb77127`
MD5	`08bb5cd581ee1a0e040ac901b86ba218`
BLAKE2b-256	`22f2a480f9d594453f92d7c13924fcb4cc0ac3e34f4a05455581e01652849f83`

Algorithm	Hash digest
SHA256	`4489e0f864e28a81d30d0825492db9b3a5c1087d0dfc1556921b0815bb210561`
MD5	`1a2482d0de57d5699ca9d030850eb8e7`
BLAKE2b-256	`1b432a48599321a216f1de6f5375ec56db74c7a67f9d1beff04db56cf4fe540a`

Algorithm	Hash digest
SHA256	`6f34b0432e56842183a9863cc01f851a52fbe2fe41c4ea47e9ee7e9d6710933b`
MD5	`294c276e5dbcefa3df2e70e85f2d2f9c`
BLAKE2b-256	`f84a0b45806b009a3950b006af0ee9e1e12af896c0e22a8183c59323925148c2`

Algorithm	Hash digest
SHA256	`9ca4d6e1ba09d5d5c90e8e79a665c0da21f8b0b4b0017a55beef42f797d437f8`
MD5	`f96c9141cee5fbb0acbc3496aa13c5ba`
BLAKE2b-256	`97d3e3fed648a05bdc786ca863c52c7e992b327c2734f288d4ab697ce9262adf`

Algorithm	Hash digest
SHA256	`b21f5812f8f4c2b46c79efc3f6320f24d575bf7007527a2a6d9c3b7e1278c27c`
MD5	`356cdf58b4aa492926ae2eae7117a09c`
BLAKE2b-256	`93e88dac78b8479a4835eed58150195aa22e25859acc2a517d08237db586d451`

Algorithm	Hash digest
SHA256	`c44e6ca04be9fc4c8e2d6ad29677062dd8bd7c2f86b48ef9ff2aa1f8b6665b8f`
MD5	`7cfe21c904d93ff3c6324add20056308`
BLAKE2b-256	`8bcfd592190eb004d9f2a37d361f1f3562946e5505f39962247e4b3e876ed717`

Algorithm	Hash digest
SHA256	`4be84dbcfd76f3f074f0372e13ddee19c9b8766ada619cabc393b5ff3ef39db5`
MD5	`05bb3405d71fd8e5677ad3a1f2be1212`
BLAKE2b-256	`ee946446ebd0758a67dea886fadf21be6c195937bbd1ed0239a1dc6c9ca179a3`

Algorithm	Hash digest
SHA256	`c09922d4326920d1ebf2a6556880b72c6822d6b7ec1ca5bf3c3c5b3aca771b6d`
MD5	`d63231e82272580fbb20f24e2ac62eee`
BLAKE2b-256	`93658b5caf44397fb71c38311904ae894ac03b0e93be40e53eb8f99034d093b3`

Algorithm	Hash digest
SHA256	`c3e33d4545a99c5ecf10e36c2579406a4360c2065dc23515fd1ae11641e1dabd`
MD5	`4fb3c71e1af942dccedb45c25c7d1246`
BLAKE2b-256	`5b8c3ecf0cb596e944d15e852d844f10e2f140a408a12b7840c15638ceb23dfe`

Algorithm	Hash digest
SHA256	`2bac10f296a370a2d0830c437f67f525b5a02342db3116028f8026738f7907a5`
MD5	`34fcb1c86b85d40d8d8b2916c0f60218`
BLAKE2b-256	`460cb557cced499b6fe5ce956ceded9bdd1e80cc514974f22e96757d49133db0`

Algorithm	Hash digest
SHA256	`5020cb86b88fac1222d594dd8ad1738c0bae358ae1d565fee79e522d1c195bc9`
MD5	`716cf1b0249d046d7d809d9c83ccf005`
BLAKE2b-256	`68772134a19afc6dc1fa81bf6357c3a0ac6904245b426a5bca5790de06e7bd0e`

Algorithm	Hash digest
SHA256	`fd850a71eebc56654bb71ae3aa8053b28ced2f29b2a9e3873c9d77663d94473b`
MD5	`1030260223373c85665dc91afc82db34`
BLAKE2b-256	`42448115a7623b34e03e418ebf8b7e799eb210309eac6b749e7eb794fc187cfa`

Algorithm	Hash digest
SHA256	`59aa6817c7b77f84d2b9cbb95efdd7939a3768a4b851bffb410149010669e788`
MD5	`c7825b488e8983cc57da523f0fb32032`
BLAKE2b-256	`cde1ec972cdd74a9988e0671f3d2491bb67944ee429dbe667804b30af06b6fd8`

Algorithm	Hash digest
SHA256	`ef6d76c98c69a74b9087a0ead600832fb293978de989337d3dbd5b3a7488f429`
MD5	`6e9fb123c4b6d6f6fdcd7b6b2a1b2166`
BLAKE2b-256	`b8d6b6cb7ee11f9f1f95b8322c1ff5f5c041ccee75321ae9e73376d783487493`

Algorithm	Hash digest
SHA256	`11a1c647b273e9f1fa181d2687ca282a3aa1f38206eecfe179b08f70a9b7d60e`
MD5	`6787eac7145473194f962378df3d0e3a`
BLAKE2b-256	`23328a66e4fe4d69f77f0feb54db42e27b67b1bb674e1265f942ec3d6b17d999`

Algorithm	Hash digest
SHA256	`41d43e45997188d9976791623ef5d79bfc3fb40249fb52a021f3d3ef5999625b`
MD5	`07848517dafaa380dbc09e830df7a559`
BLAKE2b-256	`a7a353789d3a8f47af9d353166f2a0400369ef1f84be9b8083135f72057a715b`

Algorithm	Hash digest
SHA256	`29b6f1e55e950408e0cd16d71ce8a4e1e401755bf92c89e305bd7c6e16ad2ebd`
MD5	`dca05aa6499b4fe01a2e6fd3c0ffdb74`
BLAKE2b-256	`0c9e79232091998554d3e8892b74dfd9a71e12d9b32c51b0a7bbf746322f78fa`

qoi 0.4.0

Navigation

Verified details

Maintainers

Unverified details

Project links

Meta

Classifiers

Project description

QOI

Install

Example

(Single-threaded) Benchmarks

Developing

Publishing

TODO

Discussion

Wrap or rewrite?

On the name

What's up with ./src?

USE_CYTHON=1?

Project details

Verified details

Maintainers

Unverified details

Project links

Meta

Classifiers

Release history Release notifications | RSS feed

Download files

Source Distribution

Built Distributions

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

File details

File metadata

File hashes

What's up with `./src`?

`USE_CYTHON=1`?

Algorithm	Hash digest
SHA256	`43815d72c29f51e209143d5cefde5eada850a0341e70f57f0cddc61533227036`
MD5	`032b2fbfb823ce1fd05de4284eecf2fe`
BLAKE2b-256	`d171d4f3a480b8cc7be10f89f4b1f417bdaeabba3c7f1337c042e6c3507bdd6f`

Algorithm	Hash digest
SHA256	`7f189cf1605f447b22c23d8747a4726d458b67b3237256df0ce7da0020fdec41`
MD5	`6f371609b1be1ea8a95b96547cc2ff9a`
BLAKE2b-256	`7aeb6ef3905523e7ff83c115ceb91fb66252a1865a61156522cfd60dca6be942`

Algorithm	Hash digest
SHA256	`17cca33c02811f3974c8757b7702de1e7afbf916e9bdc99d7dfdd19123799e26`
MD5	`11cfc55a8df66113ed94190ca7cd8de3`
BLAKE2b-256	`24305cc9d67b776669386dda73a3d9ae8c9df3fd2a8c27466751bdf8a153f188`

Algorithm	Hash digest
SHA256	`5ae1d935e7e8d60abcafd3951f1dd7b70a8cf165f38a825ee95af566f786cae8`
MD5	`3551eabf0c19af25d650932d0817b28c`
BLAKE2b-256	`e2e2060b3bebf114bd1001b4993250698a5ca0cabf30689148e4f165d9d92197`

Algorithm	Hash digest
SHA256	`1911239f05ab9c1b1081d80f40ea671618c1c3dd576b69486e44065c2562b7bc`
MD5	`0240d8ecb36c337fa3c83bbbd46a4013`
BLAKE2b-256	`2e5f20c2682554393d38cf5542f5358ffbd5f743ec223ea2dee40ec98996cf95`