Text Instructions for Mesh Generation?

[rrichardson]

Hi. Thank you for making and open-sourcing such an amazing tool. Setting this up and using it has been a lot of fun.

This is probably out of scope for this transformer, but I have had some occasions where the mesh generation model makes some choices that I'd prefer that it doesn't make. For example. I had it generate a mesh from a character, and it put a large, full bag on its back. (Images below) What I'd like to do is to describe the end result, so it doesn't make stylistic choices like that. Is this something that could be done, or would I have to somehow further specialize the model?

My command:

pipeline = Hunyuan3DDiTFlowMatchingPipeline.from_pretrained('tencent/Hunyuan3D-2')
mesh = pipeline(image = 'gromble1.png')

Source:

Result:

[Zeqiang-Lai]

For now, you could try to generate multiple meshes with different seed to see if the bug disappeared.

Besides, we are working on multiview input support, which could also help in this case.