Big news 🎉 We’ve just released LightlyStudio, our new open-source platform for multimodal eu1.hubs.ly/H0p04RW0 can now curate, label and manage data all in one place. 🔗 Read more: eu1.hubs.ly/H0p03lY0 💻 GitHub: eu1.hubs.ly/H0p04lm0

Select images → run plugin → bounding boxes. ⚡️ LightlyTrain brings model inference into your data workflow - no labeling scripts, no handoffs. #AutoLabeling #ObjectDetection

Did you know that LightlyStudio supports video? 🎥 Check out how to get started: eu1.hubs.ly/H0y7Q7m0

A bounding box that’s "almost right" is still wrong. 🔁 LightlyStudio’s QA tools let you catch and fix the near-misses before they corrupt your ground truth. Better labels → Better models. Always.

A biologist can spend 10 mins counting elk in a single aerial image. Colorado alone has captured more than 100,000 of them. 🦌 Vale Vision trained a model with LightlyTrain that automates the process making it way faster. Hear it from Vale Vision cofounder Joseph Porter. 🎥👇

Vale Vision is building AI for wildlife conservation. 🦌 With fewer than 100,000 images, they fine tuned DINOv3 backbones using LightlyTrain and outperformed Microsoft's MegaDetector on their own benchmarks. Read the case study: eu1.hubs.ly/H0wB90K0

YOLO has been the default for real-time object detection for years but as tasks get harder, limits start to show. Transformer based detectors address them and that is exactly why we integrated this approach into LightlyTrain. See in the full guide here: eu1.hubs.ly/H0wzLdk0

LightlyTrain covers the entire model development lifecycle. 🌐 So whether you are looking to build your own foundation model or simply want to fine-tune an existing one, LightlyTrain is built for it. Check it out here: eu1.hubs.ly/H0wyxvV0 :)

Adravision is using LightlyTrain in dentistry, and it is a great example of a pattern we keep seeing. 🦷 Check out the case study to learn more: eu1.hubs.ly/H0xdRbX0

Close to 1000 ML engineers have downloaded our guide to DINO, DINOv2, and DINOv3 since we published it. 🔍 If you want to go deep on self-supervised learning and how DINO, DINOv2, and DINOv3 fit together, the full guide is here: eu1.hubs.ly/H0wybNq0

Draw a circle in latent space ➡️ see exactly which images live there. No labels needed. No SQL filters. Just geometry. #ComputerVision #EmbeddingSpace

The other day our founders went down memory lane and dug up the very first poster of WhatToLabel, back before it became Lightly. It is a fun reminder of where this all started. Great to see how far we have come. 🚀

Found the images you needed - then lost track of them. Tags in LightlyStudio fix that. Search, pick your results, save as a tag. One click to get them back next time. No labels. No manual bookmarking. Book a demo: eu1.hubs.ly/H0wlF140

100k images and you're still poking around in file explorer? 🖼️ Grid View in LightlyStudio makes dataset exploration feel like scrolling a gallery - not grepping through folders. #CV #machinelearning

LightlyStudio is open source at its core and integrates with existing ML workflows through its Python SDK and API support. The goal is simple: help teams spend less time managing data and more time improving models. Sign up for free to test it out: eu1.hubs.ly/H0wlg4m0

🐥 People tell us our swag is cool pretty often. But hearing it is one thing. Seeing one of our users send in a cute little video starring Lightly DINO is a whole other level. If anyone else wants to show their Lightly DINO some love, we are here for it. 🦖

🦷 Adravision (YC S21) is detecting dental features on X-ray radiographs with LightlyTrain. Are you building vision models in healthcare? We'd love to help :) → Try LightlyTrain: eu1.hubs.ly/H0xs7Pc0 → GitHub: eu1.hubs.ly/H0xscRv0

LightlyStudio for UAV detection. See how our CTO Igor turns a pile of public drone datasets into something a model can actually learn from. 🚁 Full walkthrough is up on our YouTube 👉 eu1.hubs.ly/H0wlPfB0

👀 Sneak peek: Behind the scenes, our engineers have been heads down adding support for multiple embedding models in LightlyStudio. More flexibility, more ways to curate your data exactly how you want. 🚀 Stay tuned. Get started: eu1.hubs.ly/H0xgMdp0

Random sampling = biased subsets. 🧬 Diversity Sampling = coverage across your full distribution. 500 images. Every cluster represented. No wasted labels. #DataCuration #ActiveLearning #MLOps

From use case to model in three steps. 🧭 Picking the right architecture is not easy. So our team mapped the most common computer vision tasks against real deployment constraints to help you choose the right one faster. Download the cheatsheet here: eu1.hubs.ly/H0wy4Yl0