Open source

Everything we publish is available on GitHub, from the current library back to the image projects that came first.

Language models

Dlib-Transformer-extensions

Transformer architectures, language-model utilities, examples and trained models for Dlib, in C++14. Albireo and Pseudalis are developed there.

  • C++14
  • CUDA
  • BSL-1.0

ERNIE

A very small language model built with Dlib: embedding, positional encoding and linear layers written for it, with a complete training example.

  • C++
  • MIT

Dlib-BPE-Tokenizer

A byte-pair encoding tokenizer for Dlib, with vocabularies of 500 to 8,000 tokens trained on 2.7 GB of English and French text.

  • C++14
  • MIT

dlib-models-training

Training programs for Dlib models: a Vision Transformer with Barlow Twins self-supervision, a comparison with ResNet-34 on Stable ImageNet-1K, and mixture-of-experts language models.

  • C++
  • MIT

Images

Colorify

Automatic colorization of black and white pictures and films: a ResNet with a U-Net structure, refined by a generative adversarial network and trained on about 900,000 images.

Black and white pictures and their versions colorized by the high-resolution model of Colorify
  • C++
  • GAN
  • MIT

FFgan

Synthetic faces generated by a DCGAN trained on over 150,000 faces, with a built-in web server to test the model.

Synthetic faces generated by FFgan
  • C++
  • GAN
  • GPL

BAREL

A hybrid learning method that follows self-supervised learning with supervised metric learning. Its face recognition model reaches 96.1% on LFW.

  • C++
  • MIT

Data and tools

Dlib-ImageNet-Datasets

Stable ImageNet-1K datasets preprocessed at several resolutions, for quick vision experiments with Dlib.

  • C++14
  • MIT

FFspider

A multi-threaded web crawler written to build image databases for research.

  • C++
  • MIT

Models contributed to dlib-models

Several models we trained are published in the model repository of Dlib: face recognition with a DenseNet trained by the BAREL method, a gender classifier, an age predictor, synthetic faces, high-resolution colorization, a self-supervised ResNet-50 and Vision Transformer models.

Research themes on images

  • Generative adversarial networks
  • Automatic colorization
  • Noise removal
  • Tokenization
  • Self-supervised and unsupervised learning
  • Face recognition

The library and the products are under the Boost Software License 1.0. The other repositories carry their own license, shown on each card.