DinoV2: Meta’s Open Source State-of-the-art computer vision models | Hacker News Reader

DinoV2: Meta’s Open Source State-of-the-art computer vision models

ai.facebook.com

63 pointstim_sw11 comments

Well, can't believe I am saying this but I am glad we still have facebook to open sourcing their models.

We have cover texts, images, videos. I wonder what is going on with audio. Once all of these get merged into one combined model, I honestly don't see any differences between a random person on the net and an AI. The only ways we can interact with each other through the internet are those 4 things.

The key question here is: Can AI be creative? We know it can be idiotic. It can be indistinguishable from creative. But not really, or it can?

The real question is if it is sufficient to be indistinguishable from creative.

The answer is likely 'no'. As turing tests show even _stupid_ chatbots can produce impression of being humans if they are rude enough. Another example: playing chess on superhuman level has nothing to do with being smart.

These are only available for non commercial use. But still good to have.

> Creative Commons Attribution-NonCommercial 4.0 International Public License

https://github.com/facebookresearch/dinov2/blob/main/README....

Yes, DINOv2 isn't open source, but it is source-available. Hopefully a free and open source alternative can replicate its performance.

DINOv2 exists today because of the contributions of countless other maintainers in their dependency tree https://thanks.dev/e/gh/facebookresearch/dependencies if they are going to license it for non commercial, it would be nice if they actually support their community.

also I couldn't find dataset. But I guess they did mention how they made the dataset. So replicating it should not hard.

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

The segmentation is so good that one can easily see how the segments correspond to depth, with remarkable precision. Amazing work.

How can you exactly use these kind models in your application? Do you have to train a layer on top of these?

Reply on news.ycombinator.com