Frankenstein: Generating Semantic-Compositional 3D Scenes in One Tri-PlaneConference: SIGGRAPH Asia 2024 Conference Papers; December 3–6, 2024; Tokyo, JapanSIGGRAPH Asia 2024 Conference Papers (SA Conference Papers ’24), December 3–6, 2024, Tokyo, JapanDOI: 10.1145/3680528.3687672ISBN: 979-8-4007-1131-2/24/12CCS: Computing methodologies Artificial intelligenceCCS: Computing methodologies Shape modeling
Han Yan email: Note: The contributions were made during their internships at Tencent XR Vision Labs. Affiliation: Shanghai Jiao Tong University , Shanghai , China , Yang Li email: Affiliation: Tencent XR Vision Labs , Shanghai , China , Zhennan Wu email: Affiliation: The University of Tokyo , Tokyo , Japan , Shenzhou Chen email: Affiliation: Tencent XR Vision Labs , Shanghai , China , Weixuan Sun email: Affiliation: Tencent XR Vision Labs , Shanghai , China , Taizhang Shang email: Affiliation: Tencent XR Vision Labs , Shanghai , China , Weizhe Liu email: Affiliation: Tencent XR Vision Labs , Shanghai , China , Tian Chen email: Affiliation: Tencent XR Vision Labs , Shanghai , China , Xiaqiang Dai email: Affiliation: Tencent XR Vision Labs , Shanghai , China , Chao Ma email: Note: Corresponding author. Affiliation: Shanghai Jiao Tong University , Shanghai , China , Hongdong Li email: Affiliation: Australian National University , Canberra , Australia and Pan Ji email: Affiliation: Tencent XR Vision Labs , Shanghai , China
Abstract
Abstract: We present Frankenstein, a diffusion-based framework that can generate semantic-compositional 3D scenes in a single pass. Unlike existing methods that output a single, unified 3D shape, Frankenstein simultaneously generates multiple separated shapes, each corresponding to a semantically meaningful part. The 3D scene information is encoded in one single tri-plane tensor, from which multiple Signed Distance Function (SDF) fields can be decoded to represent the compositional shapes. During training, an auto-encoder compresses tri-planes into a latent space, and then the denoising diffusion process is employed to approximate the distribution of the compositional scenes. Frankenstein demonstrates promising results in generating room interiors as well as human avatars with automatically separated parts. The generated scenes facilitate many downstream applications, such as part-wise re-texturing, object rearrangement in the room or avatar cloth re-targeting. Our project page is available at:https://wolfball.github.io/frankenstein/.
原文 arXiv:2403.16210;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2403.16210v2