SeMo3D: Unifying 3D Generation, Editing, and Completion via Selective Modeling
SeMo3D trains without paired editing or completion data while preserving regions that should stay unchanged.
Training randomly selects regions of existing 3D assets and applies noise and loss only within them. Its Omni3DiT diffusion transformer separates regions to create from those to retain, combining image, text and point guidance. The authors report state-of-the-art results across three benchmarks covering generation, editing and completion, with strong consistency and accuracy on the latter two tasks.
HF Daily Papers' note
Training randomly selects regions of existing 3D assets and applies noise and loss only within them. Its Omni3DiT diffusion transformer separates regions to create from those to retain, combining image, text and point guidance. The authors report state-of-the-art results across three benchmarks covering generation, editing and completion, with strong consistency and accuracy on the latter two tasks.
HF Daily Papers' note
score 5