很不错的工作

[LaFeuilleMorte]

很棒的工作，有点像depth-anything的思路，先在高质量synthetic data上训练， 再迁移到real image上。 不过出来的模型感觉纹理和shape都比较粗糙，感觉是不是可以把triplane换掉，比如（UDF， occupancy field？），triplane的分辨率会成为模型的bottleneck。

[LaFeuilleMorte]

很棒的工作，有点像depth-anything的思路，先在高质量synthetic data上训练， 再迁移到real image上。 不过出来的模型感觉纹理和shape都比较粗糙，感觉是不是可以把triplane换掉，比如（UDF， occupancy field？），triplane的分辨率会成为模型的bottleneck。还有一点就是：
像素损失和语义是否对重建来说是充分的？有没有可能用一些几何引导（法线，深度）会更好一些

[hwjiang1510]

Hi,

Thanks for your interest in our work!

On one hand, I agree with your point on further improving the performance using better representations, better base models and geometry-aware losses. And I believe there will always some new techniques come out for improving the base model performance.

On the other hand, although they are important, they are not the priority of Real3D. In this project, we want to deliver our philosophy of scaling up models/training data by self-training on in-the-wild data, which is orthogonal to developing a better base model. We hope this philosophy can help any future models.

Best,
Hanwen