Yuan Liu

You may refer to the rendering script in the Wonder3D repo.

Hi, thanks for trying our codes. The reason for causing the problem is two-fold. First, 16 views are not enough to train a NeRF with stable renderings, especially for hash-grid-based...

SyncDreamer is designed to generate 16 views of fixed viewpoints and cannot generate other views. I'm not sure about how to generate images from other viewpoints but it indeed performs...

It seems that your tiny-cuda-nn needs to be recompiled. It is not correctly linked.

Hi, thank you for trying our codes. I use the blender 3.5. I think it would be OK to use 3.0.1 because there are no specific codes for version 3.5....

'000.png' is used as the input image. `gso-eval-meshes.zip` contains all the meshes.

You may refer to this file. https://connecthkuhk-my.sharepoint.com/:u:/g/personal/yuanly_connect_hku_hk/EXxOnqGqvGRNsDKLC7RcNmMBmfZ9Fd56ihobg1jKNnCJSg?e=9Kowh5 It may fail due to the large viewpoint difference but works in general cases.

Hi, these are the predefined poses and intrinsics in SyncDreamer. https://github.com/liuyuan-pal/SyncDreamer/blob/1fa5ecf118cbda2096978d27dd602d4d027e065d/ldm/models/diffusion/sync_dreamer.py#L274

Hi, the evaluation code has been released. You may refer to this code. COLMAP indeed fails sometimes due to sparse views so don't worry. :)

They are not in the Objaverse dataset. Some of them are from Genshin Wiki and Google image searching. It seems that I do not save the original version of them...