lightly-ai/lightly

LightlySSL is a Python framework for self-supervised learning on images. Self-supervised learning is a way to train AI vision models without needing manually labeled data. Instead of requiring a human to tag thousands of photos, the model learns useful patterns by studying the images themselves through tasks designed to build general visual understanding. The library is built on top of PyTorch, a widely used machine learning framework, and is structured to be modular. That means you can swap in your own model architecture and loss functions rather than being locked into a fixed approach. It also supports distributed training via PyTorch Lightning, which lets you spread training across multiple machines or GPUs when working with large datasets. The framework includes implementations of many established self-supervised learning methods from research, covering algorithms published between 2020 and 2024. These include approaches like BYOL, DINO, DINOv2, SimCLR, MoCo, SwAV, Barlow Twins, and about two dozen others. For each one, the repository provides runnable example code in notebooks that can be opened directly in Google Colab, a free browser-based environment, so you can try them without any local setup. The open-source library covers the core training methods. The README also mentions a commercial product from the same company with additional capabilities, including Docker support and pretraining pipelines for image classification, object detection, and segmentation. There is also a companion data platform called the Lightly Worker Solution aimed at teams processing large volumes of images. Two newer companion projects from the same team are noted in the README: LightlyTrain, which simplifies starting self-supervised or distillation pretraining in a few lines of code, and LightlyStudio, a tool for visualizing, annotating, and managing image datasets. The full README is longer than what was shown.