Questions About the Implementation of Enhancement Module.

[Jackiemin233]

Thanks for your great work!
I am trying to train the enhancement module on my dataset. However, the implementation of this module is confusing.
In my view, during training stage, the input of the ControlNet is the low-quality GT images, then the scheduler adds noise to GT images, and the diffusion UNet predicts the noise. During inference stage, the output of MV Diffusion generates the low-quality images as the input of the controlnet, and the UNet recovers the normal maps and RGB images from the Gaussian noise (The same as the pipeline figure of the Wonder3D++). However, in your released code, the generated images from the MV Diffusion are never used in the enhancement module. Instead, the input of the ControlNet is the rendered RGB images, normal maps from the textured coarse mesh, and the output of MV Diffusion model seems never used, which is different from the paper. Is there anything wrong?

Looking forward to your reply and appreciate your time!