MAISI‑3D RFlow pretrained UNet – what should I feed into class_embedding?

[arunkumar-kannan]

Hi team,
Thanks so much for the great work and the excellent codebase on MAISI model.
I’m experimenting with the MAISI‑3D RFlow pretrained UNet that ships with the model‑zoo weights. My goal is plain, unconditional generation (i.e. no ControlNet, no extra conditioning).

While inspecting the accompanying config I saw that the network was trained with num_class_embeds: 128 and therefore the checkpoint contains a class_embedding.weight of shape [128, 256].

Questions

1. What semantic meaning does this class‑embedding have in MAISI‑3D? (Does each of the 128 indices correspond to a specific dataset, organ, or sampling routine?)

2. For unconditional inference, what should I pass as class_labels? (Is feeding a tensor of zeros (torch.zeros(batch_size, dtype=torch.long)) acceptable?)

Thanks in advance for clarifying how the class_embedding is meant to be used!

[Can-Zhao]

num_class_embeds: 128 means modality categories. In the current version, we always let the input tensor to be 1. But it is an API you can use if your datasets have several categories