Awesome

Comp4D: Compositional 4D Scene Generation

The official implementation of paper "Comp4D: Compositional 4D Scene Generation".

[Project Page] | [Video (narrated)] | [Video (results)] | [Paper] | [Arxiv]

News

• 2024.8.19: Revised to support more objects.
• 2024.4.1: Released code!
• 2024.3.25: Released on arxiv!

Overview

As shown in the figure above, we introduce Compositional 4D Scene Generation. Previous works concentrate on object-centric 4D objects with limited movement. In comparison, our work extends the boundaries to the demanding task of compositional 4D scene generation. We integrate GPT-4 to decompose the scene and design proper trajectories, resulting in larger-scale movements and more realistic object interactions.

Setup

conda env create -f environment.yml
conda activate Comp4D
pip install -r requirements.txt

# 3D Gaussian Splatting modules, skip if you already installed them
# a modified gaussian splatting (+ depth, alpha rendering)
git clone --recursive https://github.com/ashawkey/diff-gaussian-rasterization
pip install ./diff-gaussian-rasterization
pip install ./simple-knn

Example Case

Prompt Case

"a butterfly flies towards the flower"

Compositional Scene training

python train_comp.py --configs arguments/comp_butterfly_flower_zs.py --expname butterflyflower_exp --cfg_override 100.0 --image_weight_override 0.02 --nn_weight 1000 --with_reg  --loss_dx_weight_override 0.005

We provide a quick overview of some important arguments:

• --expname: Experimental path.
• --configs: Configuration of scene training including prompt, object identity, object scales, and trajectory. You can also use VideoCrafter in replace of Zeroscope for video-based diffusion model.
• --image_weight: Weight of sds loss from image-based diffusion model.
• --nn_weight: Weight of knn based rigidity loss.
• --loss_dx_weight: Weight of regularization acceleration loss.

Rendering

python render_comp_video.py --skip_train --configs arguments/comp_butterfly_flower_zs.py --skip_test --model_path output_demo/date/butterflyflower_exp_date/ --iteration 3000

Static Assets Preparation

We release a set of pre-generated static assets in data/ directory. During training, we keep the static 3D Gaussians fixed and only optimize the deformation modules. We referred to the first two stages of 4D-fy to generate the static 3D objects. Then we convert them to point clouds (in data/) which are used to initialize 3D Gaussians. Thanks to the authors for sharing their awesome work!

Example case

# cd /path_to_4dfy/

## Stage 1
# python launch.py --config configs/fourdfy_stage_1_low_vram.yaml --train --gpu 0 exp_root_dir=output/ seed=0 system.prompt_processor.prompt="a flower"

## Stage 2
# ckpt=output/fourdfy_stage_1_low_vram/a_flower@timestamp/ckpts/last.ckpt
# python launch.py --config configs/fourdfy_stage_2_low_vram.yaml --train --gpu 0 exp_root_dir=output/ seed=0 system.prompt_processor.prompt="a flower" system.weights=$ckpt

## Post-Process. Convert to mesh file.
# python launch.py --config output/fourdfy_stage_2_low_vram/a_flower@timestamp/configs/parsed.yaml --export --gpu 0 \
#   resume=output/fourdfy_stage_2_low_vram/a_flower@timestamp/ckpts/last.ckpt system.exporter_type=mesh-exporter \
#   system.exporter.context_type=cuda system.exporter.fmt=obj
## saved to output/fourdfy_stage_2_low_vram/a_flower@timestamp/save/iterations-export/

## Convert to point cloud.
# cd /path_to_Comp4D/
# python mesh2ply_8w.py /path_to_4dfy/output/fourdfy_stage_2_low_vram/a_flower@timestamp/save/iterations-export/model.obj data/a_flower.ply

Acknowledgement

This work is built on many amazing research works and open-source projects. Thanks to all the authors for sharing!

• https://github.com/sherwinbahmani/4dfy
• https://github.com/hustvl/4DGaussians
• https://github.com/dreamgaussian/dreamgaussian
• https://github.com/graphdeco-inria/gaussian-splatting
• https://github.com/graphdeco-inria/diff-gaussian-rasterization
• https://github.com/threestudio-project/threestudio

Citation

If you find this repository/work helpful in your research, please consider citing the paper and starring the repo ⭐.

@article{xu2024comp4d,
  title={Comp4D: LLM-Guided Compositional 4D Scene Generation},
  author={Xu, Dejia and Liang, Hanwen and Bhatt, Neel P and Hu, Hezhen and Liang, Hanxue and Plataniotis, Konstantinos N and Wang, Zhangyang},
  journal={arXiv preprint arXiv:2403.16993},
  year={2024}
}