Pretraining & Unstable Convergence

[RuiningLi]

Awesome work and thanks for releasing the multi-view diffusion training code!

In the code, there seems to be some references about pre-training (where the shape is initialized as a sphere).

Since the paper doesn't mention pre-training at all, can you provide more details about this? In particular, did you use a pre-training stage for the reconstruction model? If so, how did you do it?

I am trying to train CRM on Google Scanned Object, a much smaller dataset than Objaverse. Even on this smaller dataset, I didn't witness very fast convergence claimed in the paper (e.g., after 20 minutes the results are far from reasonable), and it tends to be very unstable (sometimes SDF values are all positive so no mesh at all). I would hugely appreciate it if you can share some insights on why this might have happened!

Many thanks again on the awesome releases so far!