Projects
All our projects rest on one library. The products and the research programmes build on it; our earlier work on images led to it.
The library: Dlib-Transformer-extensions
Dlib-Transformer-extensions adds modern transformer modelling to Dlib, a C++ machine learning library valued for its clean interfaces. It is written in standard C++14 and published under the Boost Software License 1.0.
The project covers the whole life of a model without leaving C++, with no other dependency and no additional runtime such as Python.
Load
Read an open-weight model from a GGUF container: the llama, mistral, mixtral, gemma, qwen2, qwen3 and granite families, with quantized weights dequantized on the fly.
Run
Serve it through a shape-dynamic runtime engine, a web chat interface and an endpoint compatible with the OpenAI interface.
Convert
Generate a statically typed Dlib network whose weights are ordinary trainable parameters.
Specialize
Continue the training, fine-tune with prompt masking, or attach LoRA and DoRA adapters.
Derive
Distill a teacher into a smaller student of another architecture, or prune whole blocks and repair the rest.
Deliver
Compile the trained model into a program, or serve it from an installed service.
Scope of the library
Vision and multimodality
Vision Transformers as image backbones or as the vision tower of a model that reads text and images, trainable end to end.
Retrieval
Embedding models index a folder of documents and answer a question with the passage that matches it.
Personal data protection
Rules and small learned recognizers find personal data in five languages; format-preserving encryption replaces it reversibly.
Extended GPU memory
V²RAM streams blocks between the card, host memory and a disk store, so that a model larger than the card trains on it.
NPU and integrated GPU
Each tensor operation runs on the device measured to be fastest: the host, the Intel NPU or the integrated GPU through OpenVINO.
Processor-specific kernels
Code paths for AVX2, AVX-VNNI and AVX-512 VNNI are chosen at run time from what the processor offers.
Quantized weights on the card
CUDA kernels multiply activations with weights kept in their GGUF formats, without expanding them.
Prediction as compression
A trained network and an arithmetic coder form a lossless compressor that any processor reproduces bit for bit.
Products built on the library
Albireo and Pseudalis turn the library into two ready-to-use Windows applications: inference engines, a local service, a web application and signed updates.
- ProductsAlbireo, Pseudalis
- ServicesInference engines, local HTTPS service, web applications, signed updates
- LibraryDlib-Transformer-extensions
- FoundationDlib, standard C++14
- Hardwarex86-64 processors, Intel NPU and GPU, NVIDIA GPU
Research programmes
- GPU memory extension
- Managing the residency of tensors inside the buffer object, below the framework executor.
- Adaptive computation
- Hierarchical reasoning models trained on Sudoku-Extreme and ARC puzzles, with a halting head learned with the network.
- Pseudonymization
- Four detection devices and format-preserving encryption, measured on synthetic and hand-written texts in five languages.
- A model of the library's own
- Corpus collection, tokenizer training and measurement, pre-training, instruction tuning and benchmarking are in place for a small language model built here.
Earlier work on images
Before language models, we published image projects built on Dlib: automatic colorization, synthetic face generation, self-supervised learning and face recognition. They remain available on GitHub.
