https://invertornot.com

InvertOrNot

Make your app's dark-mode smarter with InvertOrNot. Our API predicts
if an image should be inverted for optimal dark-mode rendering.
This API is free, and is open source (both the code and the weights
of the model are available).

Example of outputs

Try it out

Upload an image to see InvertOrNot what transformation it would apply
to it.

Classify & Test Random Image!
[                    ] Classify & Transform Image!

This feature requires JavaScript to be enabled.

Original Image

Original image

Transformed Image

Transformed image

Background

Despite dark-mode's popularity, images often aren't automatically
adapted to fit it. The conservative approach is to reduce the
brightness of the image, which can degrade it significantly.
Inverting the image can be a better solution, but it can't be applied
to all images (see the poor Queen Victoria).
Using our API you can make your website more pleasant to use in
dark-mode, without having to manually adapt each image.

How does it work?

The API works by finetuning an EfficientNet model on a custom dataset
using PyTorch. By using deep learning we are able to have a much more
reliable approach to solve this problem, the only alternative being
heuristics.

How can I use it?

Documentation is available here. If possible cache or store the
results provided by the API to avoid unnecessary calls.


Endpoint Method      Input                     Output
/api/    POST   List of image   List of {invert, sha1, error} for
file            files           each file.
/api/url POST   List of image   List of {invert, sha1, error, url}
                URLs            for each URL.
/api/    POST   List of SHA-1   List of {invert, sha1, error} for
sha1            hashes          each SHA-1.


Example of a request:


                curl -X 'POST' \
                  'https://invertornot.com/api/url' \
                  -H 'accept: application/json' \
                  -H 'Content-Type: application/json' \
                  -d '[
                  "https://upload.wikimedia.org/wikipedia/commons/e/e3/Queen_Victoria_by_Bassano.jpg"
                ]'



Response:


                [
                  {
                    "invert": 0,
                    "sha1": "da487e2e9855362b4a5f4fdb531c55ae47ac6e1b",
                    "error": "",
                    "url": "https://upload.wikimedia.org/wikipedia/commons/e/e3/Queen_Victoria_by_Bassano.jpg"
                  }
                ]



If you've never adapted images for dark-mode, here's a quick example
of what the CSS can look like:

  * no inversion: filter: grayscale(50%);
  * inversion: filter: grayscale(50%) invert(100%) brightness(95%)
    hue-rotate(180deg);

Feel free to tweak these based on your dark mode setup.

Goal

InvertOrNot is a public demonstration and proof of concept of using a
NN classifier to choose how to handle images in dark-mode.

Support

InvertOrNot offers no warranties or guaranties and is done on a
best-effort basis. If you need reliable or large-scale inversion
APIs, I strongly recommend you download the FLOSS code and model and
run your own instance. The model is lightweight (16MB) and can be run
on a CPU ([?] 100ms using ONNX Runtime).

Side note

Images may be retained for training purposes. Should you prefer your
images not be saved, hosting your own instance is recommended.
Additionally, it's important to note that these images will not be
distributed; they are solely used for training purposes

Made by Mattis Megevand, code here
This idea was inspired by Gwern