Gabriele Trivigno (@gabTrivv) header
Gabriele Trivigno (@gabTrivv) profile picture

Gabriele Trivigno

@gabTrivv

Postdoc @ Meta FAIR

 Turin
Joined March 2022
  • Tweets187
  • Following399
  • Followers479
28Media
Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)Photo by Gabriele Trivigno (@gabTrivv)
If you like recipes that scale without bells and whistles, and learn robust representations from images alone, check out our latest work! Super happy to have contributed to this with such an amazing team!
Max Seitzer (@maxseitzer) profile pictureMax Seitzer@maxseitzer17h
SSL shouldn't require complex training recipes to scale, but the current SotA needs them. Introducing JEM: a principled SSL method derived from information theory. The first latent patch-level objective demonstrated from 300M→7B params. Stable training, matches or beats DINOv3.
1
20
985
Gabriele Trivigno retweeted
Claudia Cuttano (@ClaudiaCuttano) profile picture
Claudia Cuttano@ClaudiaCuttano21h
Very happy to share that I recently joined @Meta Superintelligence Labs, FAIR as a Research Scientist Intern on the Segment Anything team, based in New York 🗽 A big thanks to @kate_saenko_ and @nikhilaravi for this incredible opportunity!
Image from Claudia Cuttano (@ClaudiaCuttano)
35
9
2
400
23,741
This paper went completely under the radar! Meta researchers have developed a novel and simple self-supervised learning framework competitive with DINOv2 and DINOv3! "Scalable Patch-Level Self-Supervised Learning" The authors introduce JEM, a principled student-teacher method that learns to align patch representations across different views. Additionally, it incorporates explicit regularization to ensure representations are informative and prevent structural collapse. The framework outperforms DINOv2 in a controlled setting on dense tasks, and even beats the DINOv3 checkpoint despite it being trained on 12x less data.
Image from Tanishq Mathew Abraham, Ph.D. (@iScienceLuvr)
25
76
3
685
60,767
This week I'll be at #ECCV2026 in Malmö! A few places to catch me: - Thursday morning at the Meta booth - Friday at 11am: spotlight talk at the Meta booth on DINOv3 & CHMv2 - Friday at 4pm: presenting our “FoundYou” paper, poster #334 Come say hi if you’re around, or if you want to grab a coffee! 👋
Image from Gabriele Trivigno (@gabTrivv)
22
449
Snap a photo of your favorite t-shirt. 👕 Now imagine Segment Anything 2 could scan through thousands of images, find the exact same t-shirt, and segment it every time it shows up. We’re excited to share FoundYou: A Unified Model for Personalized Segmentation and Retrieval, accepted at #ECCV2026! 🎉
2
4
1
28
4,967
Check out FoundYou: Code: github.com/ga1i13o/FoundYou Paper: arxiv.org/abs/2608.29917v1 Try FoundYou yourself, to find your items within a db of 100M images: gmberton.github.io/demos-url…
3
313
Gabriele Trivigno retweeted
Claudia Cuttano (@ClaudiaCuttano) profile picture
Claudia Cuttano@ClaudiaCuttanoJun 15
Had so much fun at #CVPR2026! 🎉 Really happy to have presented two papers this year and discussed them with so many people. Both projects came out of my time at @visinf, TU Darmstadt, last year. Very grateful to @stefanroth for hosting me and helping shape them from the beginning, and to all my co-authors for being part of this 😊
Image from Claudia Cuttano (@ClaudiaCuttano)
4
4
96
6,122
That’s a wrap for #CVPR2026 in Denver! Despite it being my 5th CVPR, this one was a year of firsts: • first time presenting an Oral paper • first time presenting at a booth, with the DINO project at Meta • first live demo Grateful for the great conversations, feedback, and energy throughout the week!
Image from Gabriele Trivigno (@gabTrivv)
Image from Gabriele Trivigno (@gabTrivv)
2
1
49
2,801
Gabriele Trivigno retweeted
SkalskiP (@skalskip92) profile picture
SkalskiP
@skalskip92Jun 7
INSID3 segments objects across domains using ONLY ONE annotated example it works entirely without a segmentation decoder, task-specific fine-tuning, or external mask generators like SAM CVPR 2026 paper with enormous practical potential
8
80
4
623
55,656
Today! I'm going to present MARCO in Oral Session 4D at 2 pm, and later at poster #20! Don't miss it :)
Claudia Cuttano (@ClaudiaCuttano) profile pictureClaudia Cuttano@ClaudiaCuttanoMay 25
✨#CVPR2026 Oral ✨ A tale of a failed experiment: what if you fine-tune DINOv2 on sparse keypoints, beat every benchmark, only to discover it performs worse than the original frozen model on novel keypoints? 🚀MARCO closes this gap: a unified model for generalisable correspondences github.com/visinf/MARCO
Image from Claudia Cuttano (@ClaudiaCuttano)
2
8
2
52
9,876
Gabriele Trivigno retweeted
Claudia Cuttano (@ClaudiaCuttano) profile picture
Claudia Cuttano@ClaudiaCuttanoJun 6
✨ #CVPR2026 Oral ✨ INSID3 turns a frozen DINOv3 into a training-free in-context segmenter across domains and granularities! Excited to present our work today in Oral Session 4D (14:00–15:15). Come by our poster later if you’d like to chat: 📍 Poster #19 🕓 16:00–18:00 See you there!
4
19
3
126
10,738
First stroll in the poster area at @CVPR... It appears the board length is too short for the advertised poster length (the usual 2m). If the boards aren't going to be replaced, poster sessions of the main conf will be quite a mess. #CVPR2026
Image from Gabriele Trivigno (@gabTrivv)
2
6
1
48
8,124
Just landed in Denver for #CVPR2026! 🏔 Excited to present two oral papers: 📍 June 6, 14:00–15:15 — Oral Session 4D: Visual Segmentation • INSID3: Training-Free In-Context Segmentation with DINOv3 • MARCO: Navigating the Unseen Space of Semantic Correspondence Posters later that day, 16–18: • #19 — INSID3 • #20 — MARCO Also on June 4: • MARCO @ Workshop on Image Matching • INSID3 @ Workshop on Visual Concepts And don’t miss our demo: 🗓 June 5, 16–18 — FoundYou Happy to chat, come find us around! 👋
Image from Gabriele Trivigno (@gabTrivv)
2
3
27
858
MARCO: Code: github.com/visinf/MARCO Paper: arxiv.org/abs/2604.18267
2
98
INSID3: Code: github.com/visinf/INSID3 Paper: arxiv.org/abs/2603.28480
GitHub - visinf/INSID3: [CVPR 2026 Oral] "INSID3: Training-Free In-Context Segmentation with DINOv3"

GitHub - visinf/INSID3: [CVPR 2026 Oral] "INSID3: Training-Free In-Context Segmentation with DINOv3"

[CVPR 2026 Oral] "INSID3: Training-Free In-Context Segmentation with DINOv3" - visinf/INSID3

github.com
3
124
Gabriele Trivigno retweeted
Claudia Cuttano (@ClaudiaCuttano) profile picture
Claudia Cuttano@ClaudiaCuttanoMay 25
✨#CVPR2026 Oral ✨ A tale of a failed experiment: what if you fine-tune DINOv2 on sparse keypoints, beat every benchmark, only to discover it performs worse than the original frozen model on novel keypoints? 🚀MARCO closes this gap: a unified model for generalisable correspondences github.com/visinf/MARCO
Image from Claudia Cuttano (@ClaudiaCuttano)
7
57
3
448
33,917
No one is posting about it, so I feel like I should... Who's waiting on #ECCV2026 reviews? @eccvconf
2
43
7,739
What if a model could learn dense semantic matches from just a handful of annotated landmarks, while still generalizing to unseen keypoints and categories — and running 10× faster than diffusion-based approaches? MARCO is selected as an Oral at #CVPR2026! A unified model for generalizable semantic correspondence, built on DINOv2⭐️ 👉 Try our model: github.com/visinf/MARCO
4
26
2
130
13,367
more replies
3) To evaluate beyond standard benchmarks, we introduce a new benchmark with 62 novel categories from diverse domains, including animal species, furniture, and clothing, all unseen during training. We also include novel keypoint annotations on known categories, increasing coverage (e.g., from 7 to 68 landmarks for human face) This enables evaluation along two axes of generalization: to novel categories and to novel keypoints within known ones.
Image from Gabriele Trivigno (@gabTrivv)
1
2
7
578
4) MARCO is state-of-the-art, outperforming competitors on standard benchmarks. The gap further increases on unseen categories and unseen keypoints, highlighting its strong generalization. All with a single DINOv2 backbone: 3× smaller and 10× faster than diffusion-based methods.
7
381
Load more