DINOv2: State-of-the-art computer vision models with self-supervised learning | Hacker News Reader

DINOv2: State-of-the-art computer vision models with self-supervised learning

dinov2.metademolab.com

176 pointsmvoodarla16 comments

Meta AI is on fire recently. Segment Anything, Animated Drawings, and now DINOv2... all in a single month!

As a grad student in this field, Meta has always had great contributions to the open source machine learning effort, through no small effort of Yann LeCun's internal advocacy. What has changed recently is their PR strategy: OpenA(P)I has basically shown everybody that it doesn't matter if you have the best models if your publicity sucks.

I'd have to disagree. Meta was always great at releasing papers but they never scaled stuff up until like a year or so, such that things start to work in the real world outside of datasets. Glad to see, we're done with the 2% improvement in dataset papers in CVPR, I would now like to see real world performance results.

"Scaling things up" in and of itself is a very recent phenomenon. I would claim, in that sense that nothing before GPT-3 counts.

However, for not-so-scaled models, Meta or rather FAIR, had good contributions with available code that were interoperable with other research in the field. Compare that with, say, Google research/brain/deepmind, which always put out lofty papers that never allowed reproducibility.

They gave us PyTorch. That's more than enough to appreciate their contribution.

Fire and firing - you forgot the layoffs

The results are very impressive, and the model + code is made available and open source, kudos to the meta/fair team for consistently pushing out good research

I see "This repository and the models are released under the CC-BY-NC as found in the LICENSE file." at https://dinov2.metademolab.com/, have I missed something or this is definitely not FOSS?

https://mchromiak.github.io/articles/2023/Apr/18/DINOv2-Lear...

And trainable on only 96 GPUs? Loving this recent trend in vision models that are borderline tractable to replicate.

This is a progress. ;) If I'm not mistaken biggest LLaMA used 2000 GPUs for 21 day.

Even the biggest vision models have traditionally been much smaller than even medium-sized "large" language models, so I don't think the comparison is purely apple to apple.

if meta win, Nvidia win^2

>This repository and the models are released under the CC-BY-NC as found in the LICENSE file.

That's really disappointing. They keep getting so close to winning the appreciation and admiration of the community with these amazing models, but they keep squandering it.

They could at least provide a way for people/companies to pay for an unrestricted license for commercial use-cases.

If that's your opinion, you have _severely_ misunderstood Meta's endgame. As I see it, these open releases create two types of goodwill for Meta:

1. First is among the general populace, where Meta AI can be an "open" alternative to OpenAI. This helps launder the name of Meta away from all the metaverse shenanigans into something people can appreciate.

2. Second is among the academic or industry researchers that Meta may hire down the line. The association of Facebook has always made talent search hard for Meta. That can change if Meta is the "cool" place to go work.

Building a goodwill amongst the small sliver of the population who are technically competent enough to use these models but also motivated enough to create competing businesses is a non- or even anti-goal for Meta. Really, one of the worst things for them would be for Google er al to put zero effort and just being able to copy paste this work. In this case, the CC-BY-NC is doing exactly what it's supposed to do.

I think there's unsettled legal issues for companies wanting to use these.

And of course, releasing non-commercial open source software with commercial support contracts is a perfectly fine business structure, though obviously not Facebook's.

Reply on news.ycombinator.com