---
id: lightly
version: "1.5.26"
license: Copyright (c) 2018 The Python Packaging Authority Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to… (full text in the JSON record)
license_treatment: permissive
maintenance: active
---

# lightly — A deep learning package for self-supervised learning

License: permissive · Maintenance: active · Downloads: 166.3K/mo

## What it is and what it does

Lightly is a self-supervised learning framework built on PyTorch that implements established SSL algorithms for computer vision. It provides modular, low-level building blocks—loss functions, model heads, and complete model implementations—that let you train neural networks on unlabeled image data by learning visual representations through contrastive or other self-supervised objectives. The framework supports models including MoCo, SimCLR, BYOL, SwaV, DenseCL, SimSiam, Barlow Twins, DetConS, DINO, NNCLR, and LeJEPA, with examples for both single-GPU PyTorch and distributed training via PyTorch Lightning.

You use it by selecting a model architecture, defining a loss function, and training on your unlabeled dataset—the learned representations can then be fine-tuned for downstream tasks like classification or detection. The package is designed to be accessible and PyTorch-idiomatic, with extensive documentation and Colab notebooks for each model. It depends on torch, torchvision, pytorch_lightning, hydra-core for configuration, and several utility libraries, making it suitable for research and production workflows where you have unlabeled image data and want to bootstrap a strong visual encoder.

Use it for:

- Pretrain a vision encoder on a large unlabeled image dataset, then fine-tune it for classification or detection tasks downstream.
- Experiment with different SSL algorithms (MoCo, SimCLR, BYOL, DINO) to compare their effectiveness on your specific domain.
- Train distributed SSL models across multiple GPUs using PyTorch Lightning integration for large-scale pretraining.
- Build custom SSL pipelines by composing loss functions and model heads for novel self-supervised objectives.
- Reduce labeled data requirements by leveraging self-supervised pretraining before supervised fine-tuning.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Lightly provides self-supervised learning models and loss functions for computer vision, enabling you to train neural networks on unlabeled image data using methods like MoCo, SimCLR, BYOL, and DINO.

Yes. Lightly is actively maintained, MIT-licensed, and implements a mature set of self-supervised learning algorithms with low install friction. Install it if you need to train on unlabeled image data or experiment with SSL methods. The substantial dependency tree (torch, pytorch_lightning, hydra-core) is typical for deep learning and not a drawback if you're already in that ecosystem; it becomes a consideration only if you're adding it to a minimal environment.

## Install

pip install lightly
uv add lightly
poetry add lightly

## Installing lightly

Before you install: Low install friction with a pure-Python wheel. Active maintenance with a recent release (18 days old) and steady repository activity. Requires 14 runtime dependencies including torch, torchvision, and pytorch_lightning, which are substantial but standard for deep learning work.

License in practice: MIT license permits commercial and private use with minimal restrictions—you may use, modify, and distribute the package freely provided you include the license notice.

Quickstart:

pip install lightly

import lightly
from lightly.models import ResNetSimCLR
from lightly.loss import NTXentLoss

model = ResNetSimCLR()
loss = NTXentLoss()

Requires PyTorch and torchvision; GPU recommended for practical training but not strictly required.

Verify before relying:

- Whether the package includes pretrained model weights or requires training from scratch.
- Performance characteristics and typical training time for supported models on standard datasets.
- Compatibility with specific PyTorch versions beyond the stated Python version support.

## Package facts

- License: Copyright (c) 2018 The Python Packaging Authority Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to… (full text in the JSON record) (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 166.3K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

self-supervised learning computer vision, contrastive learning framework pytorch, ssl pretraining models, unlabeled image training, vision representation learning, pytorch ssl models, moco simclr byol dino, self-supervised-learning, computer-vision, pytorch

[View on SkillFed](https://skillfed.io/packages/lightly) · [View on PyPI](https://pypi.org/project/lightly/)