模型泛化性问题

[ck032]

用examples下自带的图片，生成的效果较好。用自己的图片，效果不佳。请问需要对图片做什么特别处理吗？怎么提高泛化性？

[bluestyle97]

May I ask what style of images you use? Zero123++ generalizes well to 3D-style images. It has a chance of generating flat novel views for 2D-style input images, and its ability to generate human faces is relatively poor.

[ck032]

2D/3D-style，人物/非人物的图片都有测试过。instant-nerf-large.yaml的效果比instant-mesh-large.yaml要好一些。如果想进一步提高质量，可以从哪些方面去调整呢？