Score Jacobian Chaining Toy Demo

Introduction

这个工程是对Score Jacobian Chaining（与DreamFusion相似）的简单理论实现，用作三维物体生成
This is a simple implementation of the algorithm in the Score Jacobian Chaining (The algorithm is also similar as the one of Dreamfusion), which is designed for 3D generation.

Environment

1. 至少需要安装 Pytorch, OpenCV-Python, Numpy and Tqdm 库
   Pytorch, OpenCV-Python, Numpy and Tqdm are required.

2. 需要CUDA环境
   CUDA is needed.

3. 详细的Conda环境在environment.yml中，仅供参考
   The detailed conda environment is in the environment.yml file, and for reference only.

4. 数据集可自行设计（请修改Dataset类中的加载脚本），也可以下载简单训练数据
   Simple training data can be downloaded in the Netdisk
   datasets文件夹放在本目录下即可
   Place the datasets folder in this directory
   链接(Link)：https://pan.baidu.com/s/1MjsTYkg5p2sQ1fepaBUfNQ
   提取码(Password)：b7p1

5. 运行脚本前请在本目录下新建weights文件夹用以存放网络权重
   Before Running the scripts, please create a "weights" folder in this directory

Scripts

test_chain

SJC 训练结果测试
Test Script for the result of SJC.

test_diffusion

DDPM 训练结果测试
Test Script for the result of DDPM.

test_nerf

NeRF 训练结果测试（与test_chain一样）
Test Script for the result of NeRF. The same as test_chain

train_chain

SJC方法训练NeRF（请先训练DDPM再训练SJC）
Train Script for SJC, please train DDPM first.

train_diffusion

训练DDPM
Train Script for DDPM.

train_nerf

训练NeRF（这是不使用SJC，直接用数据集图像训练NeRF）
Train Script for NeRF. This directly uses dataset images for training, rather than using SJC.