Reproduce “TEXTRIX: Latent Attribute Grid for Native Texture Generation and Beyond” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yifei Zeng, Yajie Bao, Jiachen Qian, Shuang Wu, Youtian Lin, Hao Zhu, Buyu Li, Feihu Zhang and 2 more
TEXTRIX, developed by researchers from Nanjing University, DreamTech, HKU, and OriginArk, introduces a framework that uses a latent 3D attribute grid for generating high-fidelity textures and performing high-precision 3D part segmentation from single-view inputs. The system achieves superior performance, evidenced by higher SSIM and PSNR for texture reconstruction and an improved mIoU on complex 3D meshes for segmentation compared to prior methods.
Key insights
- Operating in a native volumetric 3D space with a sparse attribute grid fundamentally avoids inter-view inconsistencies and UV seam artifacts that affect 2D-centric and UV-space texture generation methods.
- Embedding input image conditions directly into the VAE's latent space via 3D projection of a canonical view effectively bridges the domain gap, ensuring high-fidelity frontal alignment of generated attributes.
- A unified 3D attribute grid representation can serve as a versatile foundation for both generative tasks (texture, PBR material generation) and high-precision perception tasks (3D part segmentation), streamlining 3D content pipelines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.