The paper presents SeMo3D, which handles 3D generation, editing and completion by regenerating only a selected region and leaving the rest of the asset intact.
Core idea: Selective Modeling. The abstract says all three tasks synthesize a designated region while keeping everything else unchanged.
Training: Region-selective denoising applies noise and loss only to randomly chosen regions of standard 3D assets, so the abstract says it needs no paired editing or completion data.
Architecture: Omni3DiT, a multi-condition 3D diffusion transformer that separates regions to create from regions to keep, and fuses image, text and point guidance.
Results: The abstract claims it beats state-of-the-art methods on all tasks across three benchmarks, and that it can serve as a modeling tool for agents driven by natural language.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper presents SeMo3D, which handles 3D generation, editing and completion by regenerating only a selected region and leaving the rest of the asset intact.