– · Master project in Prof. Sabine Süsstrunk's lab (EPFL IVRL)

3D-BlockGen

Turned text into buildable LEGO models with a two-stage diffusion model (shape, then color) trained on 542,292 voxelized Objaverse samples. It beats single-stage variants on IoU, F1 and color.

GitHub →Dataset →Fine-tuning set → diffusion · 3D · PyTorch · CLIP

CLIP text conditioning, 32³ RGBA voxels, LEGO conversion. The 542,292 samples come from about 135,000 unique voxelized Objaverse models with 90-degree rotation augmentation (515,177 train, 27,115 test). Fine-tuning set of 11,464 samples. Code, model and data are open source.

Collaborators: Martin Everaert, Eric Bezzam.