Back to Gap Map Infrastructure / Classic ML & computer vision

Lightly

Lightly
open source / Overall score: 2.4

LightlySSL, published as the lightly package, is a PyTorch library for self-supervised pretraining of computer-vision backbones on unlabeled images. It implements about two dozen methods, among them SimCLR, MoCo, BYOL, DINO, DINOv2, MAE and iBOT, from shared losses, heads and transforms, and also trains the LeWM latent world model. Lightly, a spin-off from ETH Zurich, builds it.

Openness

5 medium confidence
5.0
license
MIT (OSI)
source
public (github.com/lightly-ai/lightly)
core features withheld
no — LightlySSL is published whole; Lightly's paid products, LightlyTrain and LightlyStudio, are separate products and withhold nothing from it

The library is MIT-licensed and builds from the public repository, with nothing held back from it. What Lightly sells, LightlyTrain for pretraining and fine-tuning and LightlyStudio for curating and labeling data, are separate products listed apart from the open LightlySSL framework.

Adoption

2 high confidence
2.0

Measured on monthly PyPI downloads of the lightly package. The package also carries the client for Lightly's hosted platform, so part of the figure may come from platform users rather than from self-supervised training.

Capability

3 medium confidence
3.0

LightlySSL is a toolkit of self-supervised training methods with recipes to pretrain and compare backbones, as timm is a zoo of architectures with training scripts. It ships no model that predicts on a user's own task without training; choosing and running a method is left to the user.

Verified 2026-11-12, except some scores, still awaiting re-confirmation