🎉 Big news: LightlyTrain now supports DINOv3. Learn more here

LightlyTrain Adds Support for DINOv3: Pretraining, Fine-Tuning, and Distillation Made Simple

LightlyTrain now supports DINOv3, bringing state-of-the-art self-supervised pretraining, distillation, and fine-tuning into a simple and flexible workflow.

Get Started with Lightly

Talk to Lightly’s computer vision team about your use case.

Book a Demo

Product

LightlyTrain

Category:

New Feature

Use Case

6 mins

We’re excited to announce that LightlyTrain now supports DINOv3, the latest generation of self-supervised vision transformers. With this update, ML teams can take advantage of DINOv3 for distillation, pretraining, and fine-tuning, all integrated into LightlyTrain’s streamlined workflow.

Just like with DINOv2, we’ve packaged DINOv3 into a simple-to-use solution that requires only a few lines of code to get started.

Pro tip

For more information about LightlyTrain, please check our documentation:

👉 LightlyTrain GitHub
👉 LightlyTrain Documentation

Distill DINOv3 Into Any Model Architecture

LightlyTrain lets you distill DINOv3 into any architecture, including YOLO, ResNet, RT-DETR, ConvNeXt, ViT, and custom models. This makes it easy to transfer the strong representation power of DINOv3 into your downstream tasks.

**Figure 1**: LightlyTrain distills DINOv3 into any model architecture.

The code snippet below shows you how to get started, head over to the LightlyTrain documentation for more information.

Getting Started: DINOv3 Distillation

Here’s how simple it is to distill DINOv3 into YOLO with LightlyTrain:

import lightly_train

if __name__ == "__main__":
    lightly_train.train(
        out="out/my_experiment",
        data="my_data_dir",
        model="ultralytics/yolo11",
        method="distillation",
        method_args={
            "teacher": "dinov3/vits16",
		 # Replace with your own url
            "teacher_url": "https://dinov3.llamameta.net/dinov3_vits16.pth"
        }
    )

‍

SOTA Fine-Tuning with DINOv3 + EoMT

We’ve also added support for EoMT, the new state-of-the-art semantic segmentation model originally built on DINOv2. By adapting it to DINOv3, LightlyTrain achieved a new SOTA on ADE20K with a ViT-L backbone:

Table 1: SOTA Fine-Tuning with DINOv3 + EoMT.
Dataset	Task	EoMT DINOv2 (ViT-L/14)	EoMT DINOv3 (ViT-L/16)
ADE20K	Semantic Segmentation mIoU	58.4	59.1

‍

Getting Started: Semantic Segmentation with DINOv3

Here’s how to fine-tune EoMT with DINOv3 using LightlyTrain:

import lightly_train

if __name__ == "__main__":
    lightly_train.train_semantic_segmentation(
        out="out/my_experiment",
        model="dinov3/vits16-eomt",
        model_args={
            "backbone_url": "https://dinov3.llamameta.net/dinov3_vits16.pth",
        },
        data={
            "train": {
                "images": "my_data_dir/train/images",
                "masks": "my_data_dir/train/masks",
            },
            "val": {
                "images": "my_data_dir/val/images",
                "masks": "my_data_dir/val/masks",
            },
            "classes": {
                0: "background",
                1: "car",
                2: "bicycle",
                # ...
            },
	 # Optional, classes that are in the dataset but should be
	 # ignored during training.
            "ignore_classes": [0],
        },
    )

‍

Try DINOv3 in Your Pipeline Today

With this release, LightlyTrain makes it easier than ever to integrate DINOv3 into real-world pipelines for detection, segmentation, and beyond. Whether you want to pretrain on your domain-specific data, distill into a custom model, or fine-tune for SOTA results, you can now do it all with just a few lines of code.

👉 Get started with LightlyTrain documentation.