Lightly

The Lightly x mimic meetup was a full house 🙌 We packed the room for a fireside chat on multimodal AI, and the conversation did not disappoint. The panel talked about how teams are actually building with multiple modalities, the data challenges that come with it, and what AI means for how we engineer software. Oh, and the pizza was great too 🍕 We're already thinking about the next one. If you missed this edition - keep an eye out! 👀 Thanks to everyone who attended!

Training computer vision models usually means facing a massive bottleneck: the need for huge labeled datasets. And let's be honest, generic models rarely fit specific, niche domains. That's why we built LightlyTrain. 🚀 We've put together a quick 1-minute overview showing how you can build production-ready CV models using pretraining, fine-tuning, autolabeling, and distillation - all in one workflow. Here's how LightlyTrain changes your workflow: ✅ Fine-Tune in a Few Lines of Code: Load a pretrained backbone and fine-tune on your own labeled data for object detection, segmentation, or classification. ✅ No Labels? No Problem: Pretrain directly on raw, unlabeled images to build stronger, domain-adapted foundations before you ever touch a label. ✅ Better Starting Point = Better Fine-Tuning: Because the backbone already understands your domain, fine-tuning converges faster and needs far fewer labeled examples to hit strong performance. ✅ Automated Annotation: Cut manual labeling work with the autolabeling and distillation pipeline. ✅ Fully On-Premises: Keep your data 100% secure and run everything locally. Getting started is as simple as one command: 💻 pip install lightly-train Check out the video below to see the full workflow in action - from pretraining on unlabeled data to fine-tuning a model ready for deployment. ▶️ Watch the full workflow in action: https://lnkd.in/e9fZgrQE 🔗 Learn more and get started: https://lnkd.in/er9PH25g

📣 Last but definitely not least, meet Igor Susmelj Co-Founder & CTO at Lightly, and one of the hosts of this Thursday's fireside chat. Igor studied Electrical Engineering at ETH Zurich and took Lightly through Y Combinator in 2021. 👀 His take on why ML teams struggle? It's rarely the models. It's the data. Fragmented workflows, hard-to-trust pipelines, datasets that fall apart at scale - that's where most teams actually get stuck. Building a product around that reality, rather than around a single clever algorithm, has been the real challenge. That kind of thinking is exactly what this evening is about. 👏 Thursday, April 23rd at 6pm, Lightly HQ Zurich. If you are a CTO/Engineer building in Switzerland, this one is for you; https://luma.com/cdnhf3bu

As summer settles in, we're excited to team up with Lightly to host a curated AI meetup at their HQ. 🗓️ Date: April 23 (Thursday) ⏲️ Time: 6pm 📍 Location: Bahnhofstrasse 86, Zürich We'll keep it real. Featuring a fireside chat with: 👉 Elvis Nava (Co-Founder & CTO at mimic) 👉 Igor Susmelj (Co-Founder & CTO at Lightly) 👉 Franziska Geiger (Senior ML Engineer at Cradle) 👉 Tiago Kieliger (Co-founder & CTO at Rivia) This is an invite-only event with limited capacity. Registration link in the comments!

🎙️ Next up in our speaker lineup for the Lightly x mimic AI Meetup - meet Tiago Kieliger, Co-Founder & CTO at Rivia. Tiago studied cybersecurity at ETH and worked for the Swiss Department of Defense before co-founding Rivia in drug development. 🔬 👉 He'll be bringing a perspective that's pretty rare in AI circles - what it actually takes to ship AI in a highly regulated industry where errors simply aren't an option. Join us this Thursday, April 23rd at 6pm at Lightly HQ in Zurich. Spots are limited: https://luma.com/cdnhf3bu

🗓️ Next week I'm co-hosting an AI meetup with our friends at mimic. We opened 60 spots. It got fully booked within 2 days. Flattering, really. But also a reminder that a *full* room isn't the same as a *good* room. I've been to enough tech events over the years to know that the thing that actually makes them worth going to is who's there. Not the agenda, not the snacks. The crowd. So we're being pretty deliberate about our waitlist. If you're an engineer/CTO building with AI, this one's for you. DM me for a spot or register here: https://luma.com/cdnhf3bu :)

We've added EUPE support to LightlyTrain 🧠 EUPE is Meta's newest visual backbone, in the same family as DINOv2 and DINOv3 - but instead of being trained purely self-supervised, it's distilled from a bunch of other Meta foundation models. EUPE might just be the strongest universal image encoder out there right now, doing really well on both global and dense tasks. With EUPE in LightlyTrain, you can: • Use it as a teacher to distill into smaller, faster models • Fine-tune it directly for things like object detection or semantic segmentation It's fully integrated, so a few lines of code are all you need to start training (snippet below 👇). Links if you want to dig in: 🔗 LightlyTrain — https://lnkd.in/esF-zV8i 🔗 EUPE — https://lnkd.in/emeAruZu 🔗 Object detection quick start — https://lnkd.in/egVCRwE9

🚀 Adding another great voice to the lineup of Lightly x mimic fireside chat: Franziska Geiger, Senior ML Engineer at Cradle. Join us on April 23rd at 6pm (next Thursday) at Lightly HQ in Zurich. Registration link: https://luma.com/cdnhf3bu More announcements soon 👀

Teaching a model to recognize dogs in a 5,000-image unlabeled dataset. Masa did it with 4 examples and a couple of refinement clicks. In this demo she shows how LightlyStudio's few-shot classifier works: use semantic search to find a few images of what you're looking for, mark them as positive examples, and let the classifier learn from there. A quick refinement loop, a few corrections, and the model knows the difference between "dog" and "not dog" across the entire dataset. 🐕 This is particularly useful when you have a large pile of raw unlabeled data and need to start organizing it fast, without setting up a full labeling pipeline first. See it in action in the video below 👇 Book a Demo to try LightlyStudio: https://lnkd.in/ewX3mPAs

🎙️ Introducing our first speaker for the Lightly x mimic AI Meetup! Meet Elvis Nava, Co-Founder & CTO at mimic. Elvis will be joining us on April 23rd at 6pm at Lightly HQ for a fireside chat on multimodal AI, system design, and building in the real world. More speaker announcements dropping soon 👀 Spots are filling up fast! RSVP here: https://luma.com/cdnhf3bu