Yuan Liu

@kxhit Hi, we follow the same codes of "fixed" to render the objects of the GSO dataset. So the poses are exactly the same as "fixed". However, I've changed the...

Hi, we simply use the image of 30°, which is the '0000.png' as input and th crop size is -1. The PSNR may be different when generating with different seeds.

I only used one instance that seems to be correct here but does not average the results. I generate 4 instances in a batch and then select the most plausible...

Sorry for the inconvenience. I only have access to OneDrive storage service provided by HKU. Maybe, it would be faster sometime. Or you may render the images from objaverse directly.

Hi, thanks for trying our codes and demo! I think it is possible to train on a single category for better performance, though we have not testes such a setting.

It's difficult because using large images will decrease the batch size and it affects the final convergence.

Hi, I only have access to infinite storage on onedrive, which is a service provided by HKU. I agree that using onedrive is sometimes annoying and I indeed have difficulty...

Hi, I've tried OpenDataLab, but it seems I cannot create a new dataset. > Hi @liuyuan-pal, thanks for your wonderful work! I wanted to express my appreciation and inquire if...

They are defined here. https://github.com/liuyuan-pal/SyncDreamer/blob/bcad0f11b72c027cc4d0464918a7f6bb07ac26e5/blender_script.py#L279

Hi, I do not try this because this requires rendering images of arbitrary viewpoints. If random viewpoints are used, the construction of the 3D feature volume should be invariant to...