Closed yuedajiong closed 8 months ago
Sorry for getting back to you late. I've just been swamped with a bunch of deadlines. (1) Indeed, the strategy of employing 2D priors for supervising 3D model training has garnered significant interest lately. But there's still a whole lot we can do with 3D data directly. Like, MeshGPT? It's got a lot of people talking. (2) Applying open-vocabulary to 3D is indeed a great idea, and we will further consider this possibility. Our work still has some flaws, and we will strive to improve them in our future work. Anyway, thanks for showing interest in our work.
很牛的工作啊。大神(们)千万不要以为我是冒犯。 我是个人技术兴趣,一直在做立体重构/生成这个方向。这个方向,稍微有点意思的论文,我都在看在学。自己也有折腾,也有自我否定,形成了自己心目中理想的算法(从需求,到技术路径),所以,看到可以借鉴的论文,都会和自己想象的理想算法去对比。然后直接的表达。 就是希望更完美,一张图一段文字就来一个sora那种级别的3D版本。
最后开个认真的的玩笑:从数据量上看:Objaverse-mix看起来3d的很大,相比于text, 2d-image做出chatgpt那种普通人觉得华丽的效果,差了至少两个数量级,咋弥补;从算力上看:8个A100训练4周,感觉是从0训练,那估计也差了两个数量级。