This repository contains the official implementation for MOVIS: Enhancing Multi-Object Novel View Synthesis for Indoor Scenes

conda create -n movis python=3.9
conda activate movis
cd MOVIS
pip install -r requirements.txt
git clone https://github.com/CompVis/taming-transformers.git
pip install -e taming-transformers/
git clone https://github.com/openai/CLIP.git
pip install -e CLIP/

Download the checkpoint and put it under MOVIS.

bash eval_single.sh

Revise the parameters within the script accordingly if one wants to change example. We use SAM and Depth-FM for getting estimated mask and depth. The background area in the depth map should be cropped out.

Download C_Obj or C3DFS_test split for benchmarking.

bash eval_batch_3d.sh
bash eval_batch_cobj.sh

You should revise the dataset path in the configs/inference_cobj.yaml and configs/inference_c3dfs.yaml file (data-params-root_dir) before running the training script.

Note that we provide the models used in C_Obj as well, if you only want to use the renderings for benchmarking, please change the path to the renderings folder.

Download image-conditioned stable diffusion checkpoint released by Lambda Labs:

wget https://cv.cs.columbia.edu/zero123/assets/sd-image-conditioned-v2.ckpt

Download the dataset from here, the dataset structure should be like this:

MOVIS-train/
    000000_004999/
        0/
        1/
        ...
    095000_099999/
    train_path.json

Run training script:

bash train.sh

One should revise the dataset path in the configs/3d_mix.yaml file (data-params-root_dir) before running the training script. Note that this training script is set for an 8-GPU system, each with 80GB of VRAM. If you have smaller GPUs, consider using smaller batch size and gradient accumulation to obtain a similar effective batch size.

This repository is based on Zero-1-to-3. We would like to thank the authors of these work for publicly releasing their code.

@article{lu2024movis,
  title={MOVIS: Enhancing Multi-Object Novel View Synthesis for Indoor Scenes},
  author={Lu, Ruijie and Chen, Yixin and Ni, Junfeng and Jia, Baoxiong and Liu, Yu and Wan, Diwen and Zeng, Gang and Huang, Siyuan},
  journal={arXiv preprint arXiv:2412.11457},
  year={2024}
}