1 Tianjin University 2 ByteDance China 3 Cardiff University
† Equal contribution * Corresponding author
This paper aims to controllably generate and edit layered 3D characters based on hand-drawing. Existing methods either generate 3D characters through global optimization or coupled representations, which makes it difficult to perform fine-grained local editing or change clothing for the character. To achieve controllable generation and part-level editing of high-quality 3D characters with changeable clothing, we propose an innovative layered 3D character generation and editing method based on sketches. Specifically, to achieve controllable generation and part-level editing in a layered manner, we propose an innovative sketch-to-3D decoupled generation network for fine-grained character generation and editing. To ensure the generated character model contains fine textures and complex geometric structures, we propose a sketch-aware progressive upsampling module to achieve high-quality character appearance. Extensive experiments on public datasets and even in-the-wild data demonstrate that our approach not only generates high-quality layered 3D characters but also enables fine-grained local editing through hand-drawn sketches.
Fig 1. The overview of our framework.
Yi Wang, Jian Ma, Zhuo Su, Guidong Wang, Yu-Kun Lai, Kun Li. "DreamCoser: Controllable Layered 3D Character Generation and Editing". SIGGRAPH Asia 2025 Technical Communications, 2025.
@inproceedings{DreamCoser,
author = {Yi Wang and Jian Ma and Zhuo Su and Guidong Wang and Yu-Kun Lai and Kun Li},
title = {DreamCoser: Controllable Layered 3D Character Generation and Editing},
booktitle = {SIGGRAPH Asia 2025 Technical Communications},
year={2025}
}