Official PyTorch implementation of paper “DreamView: Injecting View-specific Text Guidance into Text-to-3D Generation”.
This repository currently only includes the inference code for text-to-image generation (DreamView-2D). The remaining code, such as the training script and 3D generation, will be released in two months.
You can check DreamView-2D/
for the text-to-image generation code.
- 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
- Release the rendered dataset
- Release the code of DreamView-3D
- Release the training script for training DreamView-2D
- For DreamView-2D: please follow MVDream to prepare the environments
- For DreamView-3D: TBD...
We also provide our environment for reference, as shown in DreamView-2D/env-DreamView-2D.txt
- Download the checkpoints at Baidu Drive, Google Drive, ModelScope, HuggingFace
- Move the checkpoints to
ckpts/
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.
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.
TBD...
If you have any problem, feel free to open an issue or contact me :-)
- The repository of DreamView-2d is heavily based on MVDream and Stable Diffusion.
- We would like to thank the authors for sharing their exciting works.
- If you find this repo helpful, please consider citing us:
@article{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},
journal = {arXiv},
year = {2024}
}