Chat Edit 3D: Interactive 3D Scene Editing via Text Prompts ❤️ ECCV2024 ❤️

🌟 🌟 CE3D leverages an LLM to integrate dozens of different models to enable various 3D scene editing during chatting 🌟 🌟

Project Page | Paper | Video | Datasets and Ckpts |

Updates

• Hash-Atlas codes have been released in the folder of hash_atlas
• Pretrained weights and a dataset example have been released
• Release LLM-driven dialogue codes
• 💗 💗 💗 CE3D involves a large number of models and features, and there may be quite a few bugs currently. It requires the collective effort of the community! We welcome your contributions. If you have new ideas or feature implementations, please feel free to submit a pull request or contact me at any time. 💗 💗 💗

Demo (usage tutorial)

You can edit 3D scenes in the same way you interact with ChatGPT.

Quick Start

After you have set up the environment, you only need to run the following command:

# prepare your private OpenAI key.
export OPENAI_API_KEY={Your_Private_Openai_Key}

# running the demo (Advice for two Tesla A100/A800 80GB)
make run-all

# You can also run manually to select the required features and configure the GPUs.
python3 chat_edit_3D.py --port 7862 --clean_FBends --load "Segmenting_cuda:0,\
  ImageCaptioning_cuda:0,Text2Image_cuda:0,VisualQuestionAnswering_cuda:0,\
  Text2Box_cuda:0,Inpainting_cuda:0,InstructPix2Pix_cuda:0,Image2Depth_cuda:0,\
  DepthText2Image_cuda:0,SRImage_cuda:0,Image2Scribble_cuda:1,\
  ScribbleText2Image_cuda:1,Image2Canny_cuda:1,CannyText2Image_cuda:1,\
  Image2Line_cuda:1,LineText2Image_cuda:1,Image2Hed_cuda:1,HedText2Image_cuda:1,\
  Image2Pose_cuda:1,PoseText2Image_cuda:1,SegText2Image_cuda:1,\
  Image2Normal_cuda:1,NormalText2Image_cuda:1,ReferenceImageTextEditing_cuda:1"

Then visit http://localhost:7862 on the web to freely edit the 3D scene.

The above operation integrates more than 20 visual models and requires about 100G of GPU memory. If you don't have it, you can run the small version of CE3D, which only requires 20G of GPU memory (but the editing ability is limited). The commands are as follows:

make run-small-instruct

# You can also run manually:
python3 chat_edit_3D.py --port 7862 --clean_FBends --load "Segmenting_cuda:0,\
ImageCaptioning_cuda:0,VisualQuestionAnswering_cuda:0,Text2Box_cuda:0,\
Inpainting_cuda:0,InstructPix2Pix_cuda:0"

Step 1: Create a new environment & Activate the new environment

https://github.com/Fangkang515/CE3D.git
cd CE3D

conda create -n CE3D python=3.10
conda activate CE3D

pip install -r requirements.txt
pip install  git+https://github.com/IDEA-Research/GroundingDINO.git
pip install  git+https://github.com/facebookresearch/segment-anything.git

install tiny-cuda-nn according to https://github.com/NVlabs/tiny-cuda-nn/tree/master

Step 2: Download the checkpoints / pretrained-weights / dataset examples

# optional: Download the checkpoints of Sam, Dino, et al. (Directly running the code will also automatically download this file)
unzip checkpoints.zip

# Download the pretrained-weights (atlas)
cd hash_atlas
unzip weights.zip
cd ..

# Download the dataset example
cd datasets
unzip flower.zip
cd ..

Tips: If you want to train the Atlas manually

# note: (the final code will be automatically executed to obtain Atlas of a 3D scene)
# 
cd hash_atlas
python atlas_playground.py

# Then You will find an atlas_space folder in the ../datasets/flower directory, like below:
 |-flower   
    |-XXX
    |-work_space
    |  |atlas_space
    |  |  |flow
    |  |  |mask
    |  |  |propainter
    |  |  |training_results
    |  |final_atlas
    |  |  | atlas_ori_foreground.png
    |  |  | atlas_ori_background.png
    |  |editing_space
    |  |  |XXX

Citation

If you find our code or paper useful, please consider citing.

@article{fang2024chat,
  title={Chat-Edit-3D: Interactive 3D Scene Editing via Text Prompts},
  author={Fang, Shuangkang and Wang, Yufeng and Tsai, Yi-Hsuan and Yang, Yi and Ding, Wenrui and Zhou, Shuchang and Yang, Ming-Hsuan},
  journal={arXiv preprint arXiv:2407.06842},
  year={2024}
}