Questions about self.label_emb

[2hiTee]

Thanks for your wonderful work! I have a question that you said you based on pre-trained SVD to train the first stage Hi3D model. But the original SVD used FPS and bucket_id as the additional condition together with timestep embedding, and the dimension is 768. Here in the configuration file, I see you changed these two condition with elevation and aesthetic condition but starting with the same label embedding. Do you think this works well? Thanks!

[yanghb22-fdu]

In our experiments, the FPS and bucket_id conditions were found to be insignificant, so we replaced these conditions. For the new setup, elevation as a 3D condition has a significant impact on the final results. To simplify implementation and minimize changes to the network architecture and code, we opted for an aesthetic condition.