Is there any data augmentation during the training of slat flow matching models?

[LiaoLW]

Thank you for your work!
When I'm testing the pretrained slat flow matching models, I found that this model is super robust to the generated coords (from the ss flow matching model). No matter the generated coords are with many false voxels/ small scaled geometry, it will accurately identify the correct subset and generate coherent slat feature for these voxels.
But I cannot find any scaling/ noise injection within the data process pipeline or dataset, which means the model have only seen the 3D assets in the normalized scale with clean coords, so I don't understand how does the slat flow matching model achieves such robustness.
The following pictures are original generated result from trellis/ coords visualization without slat feature(demonstrating the noise coords)/ rescaled coords before stage 2 generation.