OmniPart generates part-aware, controllable, and high-quality 3D content through two key stages: part structure planning and structured part latent generation. Built upon TRELLIS, which provides a spatially structured sparse voxel latent space, OmniPart first predicts part-level bounding boxes via an auto-regressive planner. Then, part-specific latent codes are generated through fine-tuning of a large-scale shape model pretrained on overall objects.

Repository on Huggingface

Project page

Running:

1. First, upload your image, but set “generate_3d”: “false” and run to see the segmentation results.

2. Provide merging in the format ‘0,1;3,4’ (merges segments 0&1 together and 3&4 together).

You need to merge in sequential order, for example, for the sequence 3,10;3,2;9,4;0,6;9,0;8,5;1,7;1,11;8,1: 10 is merged into 3 and no longer exists, then 2 is merged into 3. 4 is merged into 9, and no longer exists - there’s only 9. Then 0 is merged into 9, etc.

Refer to the merged image before enabling generate_3d . There can be error sometimes when it does not know how to merge - see logs.

1. After the merge is ok, run full pipeline with “generate_3d”: “true”.

2. Please tweak the viewer settings to see the actual mesh - decrease brightness or increase contrast!