Lightly

Building computer vision models at scale is still harder than it should be. Too much unlabeled data sitting unused. Too many annotation bottlenecks. Too much time lost before a model ever reaches production. We built LightlyTrain and LightlyStudio to fix exactly that - together, they cover the full end-to-end CV pipeline: 🔹 LightlyTrain - Self-supervised pretraining for vision models, no labels required. Works with ResNet, YOLO, Vision Transformers, and more. Achieve up to 36% higher mAP compared to standard supervised approaches, in a fraction of the time and cost. 🔹 LightlyStudio - A unified data platform for curation, annotation, QA, and dataset management. Built for ML engineers and labelers alike, with Python SDK, open-source foundations, and enterprise-grade security. Together: curate your data in Studio → pretrain on it with Train → fine-tune and deploy. No label dependency to get started. Both products are open-source at their core and free to try. ⭐ GitHub: https://lnkd.in/eSpuzJRz 🌐 Learn more: https://lnkd.in/eB8cbfSJ

Hard not to feel proud looking at this chart. We launched LightlyTrain publicly and started tracking Weekly Active Users in March. What you see is what happened next, entirely driven by the open-source community finding value in what we built. It's a reminder of why we've always been open-source first. We did it with LightlySSL, we're doing it again with LightlyTrain and LightlyStudio. Contributing to a community that gives back so much feels like the only way to build. Big thank you to everyone who's starred the repo, filed issues, and pushed us to keep improving. You're the reason this line keeps going up. Curious what we're building? Check it out: https://lnkd.in/e8cZzWes Github: https://lnkd.in/eH3pnFsc

When you're working with large datasets, being able to quickly isolate a specific slice saves a lot of time. In this short demo, Masa shows how to filter by annotation labels and by metadata in LightlyStudio. Select a label like "airplane" and the view updates instantly - or pick multiple labels at once. And if your images have metadata attached, whether that's width, height, or any custom fields, you can easily filter by that too. 🎯 See it in action in the video below. Book a Demo to try LightlyStudio: https://lnkd.in/eCH765jh

That's a wrap on #CVPR2026 ✅ Three days of incredible conversations, demos and ideas with the computer vision community. Grateful to everyone who came by the booth! Proud of the team for making it happen. Until next year. 🙌

☀️ Day 2 of main conference at #CVPR2026 and the energy at our booth has been incredible. Nonstop conversations with researchers and engineers about training, fine-tuning, and deploying vision models at scale. Huge thanks to everyone who stopped by to chat, demo, and swap notes. 😍

Day 1 of the main conference at #CVPR2026 🙌 The booth hasn't been quiet for a minute. Yesterday we said curation is the bottleneck and today the line at the booth said it back to us 😁 So many great chats about LightlyStudio and LightlyTrain - thanks for the energy, Denver. Back at it tomorrow 👋

What a week at ICRA 2026 🤖🤯 Our CTO Igor Susmelj spent the week connecting with engineers and robotics startups from across the field. Great conversations about where autonomy and perception are heading and what it actually takes to make systems work outside the lab. A highlight was the Agricultural Robotics Workshop where we presented our collaboration with Blue River Technology (John Deere) as an oral spotlight and poster. The room was packed and the conversations were exactly the kind we love: deep, specific, and full of people who are actually building things 🚜 Exciting space to be in right now. See you at the next one 👋

Two days of workshops and paper sessions at #CVPR26 are wrapped, and what a couple of days it was. So many familiar faces, a notebook full of ideas, and a brain that’s pleasantly fried🤯 Today the main conference starts and I can’t wait! Come find us at the Lightly Booth, I’d love to chat and hear what you’re building. P.S: Get there early, the swag is pretty nice and it tends to go fast🐥