正在加载视频...
视频加载失败
Meta released SAM3 a model to reconstruct objects from segmented images. I tried it and here are 4 things I found out - The generation is incredibly fast when compared with other 3D GenAI models - The models generated are not meshes but Gaussian Splatting. I was surprised that... show more
32,447 次观看 • 10 个月前 •via X (Twitter)
9 条评论

Thanks for trying out the model! In the demo we supported Gaussian Splatting, but the model can output mesh in our open-source release here:

You can download the ply files via the browser console's network tab. Just grab the biggest files and rename

Ah nice

@CreatedByJannn

@artist_jodi

Very good for visualising potential furniture configurations. Very bad for avoiding stubbed toes :).

Is it using GS? From my experiments with it, looks like it's using some other generative modeling. I would give the same image multiple times and click on the same objects and each time I would get something different. Some were close, but most looked completely incorrect as in it generated completely different geometries each time.

And that's why it's completely useless. wake me when it makes geometry with correct topology and uvs

do you happen to know any splatting wizards who ciuld lend me their brains for a bit? got some ideas for how to make this kind of tech useful for games and situations where you dont have a photo to start from but i don't have the knowhow to know if its viable
