The conservative solution to adapt images for dark-mode consist in dimming the image, however there is a lot of images that can be inverted (graph for example). Using deep learning we can avoid heuristics and obtain a much more reliable solution.

The API uses an EfficientNet pre-trained model fine-tuned on a custom dataset (1.1k samples). EfficientNet was chosen as it was pre-trained and offered the best performance for its size.

The trained model is very small (16MB) which means you can easily run your own instance. This problem is very simple for deep learning as it's a simple binary classification.

For this project training the model wasn't the challenge as most of the time was spent on the construction of the dataset.

For the API I'm using FastAPI, Redis and ONNX Runtime to run the model. The API can be used by posting the images to the API, using URL and using SHA-1 for already processed images.

The API is free and open-sourced (http://github.com/mattismegevand/invertornot).