Skip to content

Latest commit

 

History

History
174 lines (152 loc) · 6.27 KB

README.md

File metadata and controls

174 lines (152 loc) · 6.27 KB
Raw

DreamComposer: Controllable 3D Object Generation via Multi-View Conditions

Project Page | Paper

Conda Environment

Install packages in requirements.txt. We test our model on a 24G RTX4090 GPU with Python 3.10, CUDA 11.8 and Pytorch 2.0.1.

conda create -n dreamcomposer
conda activate dreamcomposer
pip install -r requirements.txt
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Preparation for inference

  1. Download checkpoints here. (Here is a simple tutorial for downloading data from ONEDRIVE using command line.)
  2. Remove the background of images with Clipdrop.

Inference

Parameters Explanation:

  • --ckpt is the checkpoint to load.
  • --output is the output directory.
  • --sample_num is the number of instances to generate. sample_num 4 means generating 4 times.
  • --cfg_scale is the classifier-free-guidance.
  • --elevation is the elevation angle of the input image in degree. As shown in the following figure. (Don't need a very accurate elevation, but a rough value in [0,40] is OK, e.g. {0,15,30}.)
  • --crop_size Resize the input image to the crop_size, and then put it into an 256*256. crop_size=-1 means do not resize the object but only directly resize the input image to 256*256.

DC-Zero-1-to-3

  1. Make sure you have the following models.
DreamCompoer
|-- ckpt
    |-- DC_zero123.ckpt
  1. Set the input images paths and azimuths at # prepare data in dc_zero123_nvs.py.
  2. Run DreamCompoer to produce controllable novel views with a better local control.
python dc_zero123_nvs.py --ckpt ckpt/DC_zero123.ckpt \
                   --output output_imgs/dc_zero123/alarm \
                   --target_azim 45 \
                   --sample_num 4 \
                   --cfg_scale 2.0 \
                   --elevation 30 \
                   --crop_size -1 \
                   --seed 6033

DC-SyncDreamder

  1. Make sure you have the following models.
DreamCompoer
|-- ckpt
    |-- ViT-L-14.ckpt
    |-- DC_syncdreamer.ckpt
  1. Set the input images paths and azimuths at # prepare data in dc_syncdreamer_nvs.py.
  2. Run DreamCompoer to produce controllable and consistent novel views.
python dc_syncdreamer_nvs.py --ckpt ckpt/DC_syncdreamer.ckpt \
                   --output output_imgs/dc_syncdreamer/alarm \
                   --sample_num 4 \
                   --cfg_scale 2.0 \
                   --elevation 30 \
                   --crop_size -1 \
                   --seed 6033
  1. Following SyncDreamer, we use NeuS for 3D reconstruction to generate videos and meshes.
python train_renderer.py -i output_imgs/alarm/0.png \
                         -n alarm-neus \
                         -b configs/neus.yaml \
                         -l output/renderer

Parameters Explanation:

  • -i contains the multiview images generated by DC-SyncDreamer.
  • -n means the name.
  • -l means the log dir. Results will be saved to <log_dir>/<name> i.e. output/renderer/alarm-neus.

Preparation for training

  1. Generate multi-view renderings or download public rendering data. Need rendered images such as camera elevation unchanged and azimuth changing from 0° to 360°. (e.g. SyncDreamer, Gobjaverse)
  2. Modify the dataset in ldm/data/objaverse.py.
  3. Donwload the pretrained zero123-xl model here.
  4. Donwload the pretrained syncdreamer model here.

Training

  1. Pre-train the 3D lifting module
python dc_zero123_training.py \
    -t \
    -n lifting_3d \
    --base configs/DC_3d_lifting.yaml \
    --gpus 0,1,2,3,4,5,6,7 \
    --scale_lr False \
    --num_nodes 1 \
    --seed 42
  1. Train DC_Zero123

Make sure you have the following models:

DreamCompoer
|-- ckpt
    |-- zero123-xl.ckpt
    |-- lifting_3d.ckpt

Run:

python dc_zero123_training.py \
    -t \
    -n dc_zero123 \
    --base configs/DC_zero123.yaml \
    --gpus 0,1,2,3,4,5,6,7 \
    --scale_lr False \
    --num_nodes 1 \
    --seed 42 \
    --finetune_from ckpt/zero123-xl.ckpt \
    --lift3d_ckpt ckpt/lifting_3d.ckpt
  1. Train DC_SyncDreamer

Make sure you have the following models:

DreamCompoer
|-- ckpt
    |-- ViT-L-14.ckpt
    |-- syncdreamer-pretrain.ckpt
    |-- lifting_3d.ckpt

Run:

python dc_syncdreamer_training.py \
    -t \
    -n dc_syncdreamer \
    --base configs/DC_syncdreamer.yaml \
    --gpus 0,1,2,3,4,5,6,7 \
    --scale_lr False \
    --num_nodes 1 \
    --seed 42 \
    --finetune_from ckpt/syncdreamer-pretrain.ckpt \
    --lift3d_ckpt ckpt/lifting_3d.ckpt

Acknowledgement

DreamComposer is inspired by the following repos: zero123, SyncDreamer, threestudio, camtools, eg3d, ControlNet, NeuS.

Many thanks to the authors for sharing their codes.

Citation

If you find DreamComposer useful in your project, please cite our work. :)

@article{yang2023dreamcomposer,
    title={DreamComposer: Controllable 3D Object Generation via Multi-View Conditions},
    author={Yang, Yunhan and Huang, Yukun and Wu, Xiaoyang and Guo, Yuan-Chen and Zhang, Song-Hai and Zhao, Hengshuang and He, Tong and Liu, Xihui},
    journal={arXiv preprint arXiv:2312.03611},
    year={2023}
}