Sherpa3D: Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior

Fangfu Liu¹, Diankun Wu¹, Yi Wei¹, Yongming Rao², Yueqi Duan¹

¹Tsinghua University, ²BAAI

CVPR (2024)

Paper | ArXiv | Project Page | Code

Abstract: Recently, 3D content creation from text prompts has demonstrated remarkable progress by utilizing 2D and 3D diffusion models. While 3D diffusion models ensure great multi-view consistency, their ability to generate high-quality and diverse 3D assets is hindered by the limited 3D data. In contrast, 2D diffusion models find a distillation approach that achieves excellent generalization and rich details without any 3D data. However, 2D lifting methods suffer from inherent view-agnostic ambiguity thereby leading to serious multi-face Janus issues, where text prompts fail to provide sufficient guidance to learn coherent 3D results. Instead of retraining a costly viewpoint-aware model, we study how to fully exploit easily accessible coarse 3D knowledge to enhance the prompts and guide 2D lifting optimization for refinement. In this paper, we propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric consistency simultaneously. Specifically, we design a pair of guiding strategies derived from the coarse 3D prior generated by the 3D diffusion model: a structural guidance for geometric fidelity and a semantic guidance for 3D coherence. Employing the two types of guidance, the 2D diffusion model enriches the 3D content with diversified and high-quality results. Extensive experiments show the superiority of our Sherpa3D over the state-of-the-art text-to-3D methods in terms of quality and 3D consistency.

Table of Contents

• Table of Contents
• Installation
• Training
  • Quick Start
  • Custom training
• Acknowledgement
• BibTex

Installation

1. Clone Sherpa3D.

git clone https://github.com/liuff19/Sherpa3D.git
cd Sherpa3D

2. Create the environment, here we show an example using conda.

conda create -n Sherpa3D python=3.9
conda activate Sherpa3D 

3. Install PyTorch >= 1.12. We have tested on torch1.12.1+cu113.

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113

4. Install dependencies:

pip install -r requirements.txt

5. Download finetuned shap-E model.

cd load/shapE-finetuned/
bash download.sh

6. (Optional) If you are experiencing unstable connections with Hugging Face, we suggest you downloading the guidance model you used to a local folder following here and here, and set model_key of the guidance and the prompt processor to the local path.

Training

Quick Start

We provide some scripts and pre-defined configs to automatically run the geometry modeling stage and appearance modeling stage. For a quick start, you can run:

# Prompt: A head of Terracotta Army
cd bash
bash terracotta_army_1gpu.sh 

# Prompt: A detailed and realistic 3D model of a vintage camera
cd bash
bash camera_4gpu.sh

We also provide scripts for some other prompts, which you can find in the "./bash" directory. The outputs will be saved in "./outputs/geometry" and "./outputs/appearance" respectively.

Custom training

1. To perform custom training, you need to create a config file first. We have provided some examples in the "./configs" directory.
2. After completing the configuration, you can start custom training. We provide the following two methods:

• If you want to complete the training of both geometry and appearance in one go, please make sure your config files are named following the formats "xxx_geo.json" and "xxx_app.json", respectively. Then use the following commands to train:

cd bash
# using 1 GPU
general_1gpu xxx 
# using 4 GPU
general_4gpu xxx 
# using 8 GPU
general_8gpu xxx 

• If you want to train geometry and appearance respectively. You can use the following commands:

# Multi-GPU training (recommanded)
# Geometry modeling using 8 GPU 
python3 -m torch.distributed.launch --nproc_per_node=8 train.py --config path/to/your/config_geo/file
# Geometry modeling using 4 GPU
python3 -m torch.distributed.launch --nproc_per_node=4 train.py --config path/to/your/config_geo/file
# Appearance modeling using 8 GPU
python3 -m torch.distributed.launch --nproc_per_node=8 train.py --config path/to/your/config_app/file
# Appearance modeling using 4 GPU
python3 -m torch.distributed.launch --nproc_per_node=4 train.py --config path/to/your/config_app/file

# Single-GPU training
# Geometry modeling. 
python3 train.py --config path/to/your/config_geo/file
# Appearance modeling.
python3  train.py --config path/to/your/config_geo/file

Acknowledgement

This repo benefits from Fantasia3D, Threestudio, Shap-E and Cap3d. Thanks for their wonderful work.

BibTex

@misc{liu2023sherpa3d,
      title={Sherpa3D: Boosting High-Fidelity Text-to-3D Generation via Coarse 3D Prior}, 
      author={Fangfu Liu and Diankun Wu and Yi Wei and Yongming Rao and Yueqi Duan},
      year={2023},
      eprint={2312.06655},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}