Skip to content

MAISI missing labels for conditioning maisi-rflow with vista3D pseudo labels #2047

@cugwu

Description

@cugwu

I’m training a pre-trained maisi-rflow model to generate new images conditioned on segmentation masks generated with vista3D.

In MAISI-data is written that the pseudo-labels can be produced using VISTA3D. I’m using MONAI bundle. Both MAISI and VISTA3D have labels IDs from 1 to 132 as can be compared in this two links:
1.https://github.com/Project-MONAI/tutorials/blob/main/generation/maisi/configs/label_dict.json
2. https://github.com/Project-MONAI/VISTA/blob/main/vista3d/data/jsons/label_dict.json
However, when I follow the MAISI inference tutorial and download the provided pseudo-labels, I see the label 200=body which does not exist in VISTA3D. As a result, when I condition maisi-rflow on my masks that only include IDs up to 132, generation quality degrades due to the missing label.

Which model or pipeline should I use to produce complete pseudo-labels compatible with maisi-rflow conditioning? Below are examples of generated images.

Image

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions