Multi view inference

[Alex-experiments]

Hi and thanks for your excellent paper! I was wondering if it would be possible to replace the multi-view generation from MVDream with other images from the same subject.

My goal would be to take 4 images of an object, obtain camera pose estimation with COLMAP (or any other method), and do the inference with it.

Would it be feasible? I know that during training the camera poses are already randomized, but you're still keeping a front-facing image, so I was wondering if not having one would hinder performances.

[al3dv2]

I am interested by the same thing. It would be great if we could use our own images of 4 angles.