MeshFIM: Local Low-Poly Mesh Editing via Fill-in-the-Middle Autoregressive Generation
Researchers have introduced MeshFIM, a novel framework designed for local editing of low-poly meshes using Fill-in-the-Middle (FIM) autoregressive generation. Traditional autoregressive models require regenerating entire meshes when local regions are unsatisfactory, leading to computational waste and loss of existing structure. MeshFIM addresses this by regenerating only target regions based on surrounding context. The framework tackles three specific challenges: ensuring exact boundary attachment, preserving topological order, and preventing overflow beyond intended areas. It employs five key design choices, including boundary vertex markers, context positional embeddings, expanded context width, context augmentation, and a low-poly geometry encoder with a gated subtraction mechanism. This encoder focuses generation on missing regions by leveraging differences between reference surfaces and existing meshes. Ablation studies confirm the effectiveness of each component. The technology enables applications such as interactive brush-based editing and automatic defect repair. Experimental results demonstrate that MeshFIM outperforms various baselines in mesh refinement, repair, and whole mesh generation combined with stitch-back schemes, marking a significant advancement in computer graphics and AI-driven 3D modeling efficiency.
Wire timeline
MeshFIM: Local Low-Poly Mesh Editing via Fill-in-the-Middle Autoregressive Generation
Researchers have introduced MeshFIM, a novel framework designed for local editing of low-poly meshes using Fill-in-the-Middle (FIM) autoregressive generation. Traditional autoregressive models require regenerating entire meshes when local regions are unsatisfactory, leading to computational waste and loss of existing structure. MeshFIM addresses this by regenerating only target regions based on surrounding context. The framework tackles three specific challenges: ensuring exact boundary attachment, preserving topological order, and preventing overflow beyond intended areas. It employs five key design choices, including boundary vertex markers, context positional embeddings, expanded context width, context augmentation, and a low-poly geometry encoder with a gated subtraction mechanism. This encoder focuses generation on missing regions by leveraging differences between reference surfaces and existing meshes. Ablation studies confirm the effectiveness of each component. The technology enables applications such as interactive brush-based editing and automatic defect repair. Experimental results demonstrate that MeshFIM outperforms various baselines in mesh refinement, repair, and whole mesh generation combined with stitch-back schemes, marking a significant advancement in computer graphics and AI-driven 3D modeling efficiency.
cs.AI updates on arXiv.org