Awesome

DreamView

Official PyTorch implementation of ECCV 2024 paper “DreamView: Injecting View-specific Text Guidance into Text-to-3D Generation”.

This repository includes the training and inference code for text-to-image generation (DreamView-2D), and text-to-3D generation (DreamView-3D).

You can check DreamView-2D/ and DreamView-3D/ for the text-to-image and text-to-3D generation codes, respectively.

News

• 2024/7/22: We release the code and script for training DreamView-2D and text-to-3D generation via DreamView-3D
• 2024/7/3: Our paper is accepted by ECCV 2024, congratulations and many thanks to the co-authors!🥳🎉🎊
• 2024/4/11: We release the inference code of DreamView-2D, and the DreamView checkpoint: Baidu Drive, Google Drive, ModelScope, HuggingFace
• 2024/4/10: Our arxiv paper can be found here

TODO

• Release the code of DreamView-3D
• Release the training script for training DreamView-2D

Preliminary

Environment

• For DreamView-2D: please follow MVDream to prepare the environments
• For DreamView-3D: please follow threestudio to prepare the environments

Checkpoint

• Download the checkpoints at Baidu Drive, Google Drive, ModelScope, HuggingFace
• Move the checkpoints to ckpts/

Text-to-image Generation

Inference

Running the below script

cd DreamView-2D
python t2i.py --num_samples 4 --fp16

and you are expected to obtain the below result: It takes about 10G GPU memory to run the text-to-image generation, and you can modify the DreamView-2D/t2i.py to generate your own content.

Training

cd DreamView-2D
bash train.sh

Note that we use 4 8*V100 machine to train DreamView-2D by default, and to accelerate convergence, you can consider using MVDream as the initialization parameter.

Text-to-3D Generation

Running the script below to reproduce the results shown above

cd DreamView-3D
bash reproduce.sh

Note that the above script may require ~60G GPU memory, so you may run it with an A100 GPU.

Acknowledgement

• The code of DreamView-2d is heavily based on MVDream and Stable Diffusion.
• The code of DreamView-3d is heavily based on MVDream-threestudio and threestudio.
• We would like to thank the authors for sharing their exciting works.

Citation

• If you find this repo helpful, please consider citing us:

@inproceedings{yan2024DreamView,
  author = {Yan, Junkai and Gao, Yipeng and Yang, Qize and Wei, Xihan and Xie, Xuansong and Wu, Ancong and Zheng, Wei-Shi},
  title = {DreamView: Injecting View-specific Text Guidance into Text-to-3D Generation},
  booktitle = {ECCV},
  year = {2024}
}