DreamPropeller: Supercharge Text-to-3D Generation with Parallel Sampling (CVPR 2024 Highlight)

Official Implementation of DreamPropeller: Supercharge Text-to-3D Generation with Parallel Sampling

Linqi Zhou¹, Andy Shih¹, Chenlin Meng^1,2, Stefano Ermon¹

¹Stanford University, ²Pika Labs

[Paper] [Project]

---

This codebase is based on threestudio with minor changes to some files for parallel processing.

Installation

• Install Python >= 3.11.
• Create a virtual environment:

conda create -n dreampropeller python=3.11

• (Optional, Recommended) Install ninja to speed up the compilation of CUDA extensions:

pip install ninja

• Install dependencies:

pip install -r requirements.txt

• Install PyTorch == 2.1.0. We have tested on torch2.1.0+cu121.

pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121

• (Optional, Recommended) The best-performing models in threestudio use the newly-released T2I model DeepFloyd IF, which currently requires signing a license agreement. If you would like to use these models, you need to accept the license on the model card of DeepFloyd IF, and login into the Hugging Face hub in the terminal by huggingface-cli login.

Usage

We can compare the runtime for baseline and DreamPropeller by using launch_baseline.py and launch_speedup.py respectively. For DreamPropeller, one can tune speedup parameters in .yaml config files. The preset values in the files are the default values. We note their meaning below:

• speedup.threshold denotes the threshold for fixed-point error. Larger threshold gives greater speedup but lower generation quality.
• speedup.ema_decay denotes the EMA rate of updating the threshold. This parameter makes it robust against variation in speedup for different prompts.
• speedup.P denotes the window size, typically set to num_gpu-1. e.g. 8 GPU cluster uses window size of 7.
• speedup.adaptivity_type denotes the error aggregation function used to update speedup.ema_decay. We only support median and mean and we use median for all but ProlificDreamer. We find using median makes quality more stable for tested frameworks but gives slower speedup than mean.

Other visualization options:

• trainer.visualize_progress sets whether to visualize the generation progress.
• trainer.display_time sets whether to display current runtime for each displayed image.
• train_config.val_check_interval denotes the wallclock time in seconds (instead of iteration number) after which the script will save a visualization.

We can quickly test each of the frameworks presented in the paper below. Our method is tested on an 8-GPU A100-80G PCIe cluster.

You can freely switch between baseline and DreamPropeller by simply setting FILE to launch_baseline.py or launch_speedup.py.

DreamFusion

python $FILE --config configs/dreamfusion-if.yaml --train system.prompt_processor.prompt="an ice cream sundae" data.batch_size=16 trainer.visualize_progress=true

Magic3D

python $FILE  --config configs/magic3d-coarse-if.yaml --train system.prompt_processor.prompt="a beautiful dress made out of fruit, on a mannequin. Studio lighting, high quality, high resolution" data.batch_size=16 trainer.visualize_progress=true

python $FILE --config configs/magic3d-refine-sd.yaml --train system.prompt_processor.prompt="a beautiful dress made out of fruit, on a mannequin. Studio lighting, high quality, high resolution" data.batch_size=16 trainer.visualize_progress=true system.geometry_convert_from=path/to/coarse/stage/trial/dir/ckpts/last.ckpt

TextMesh

python $FILE --config configs/textmesh-if.yaml --train system.prompt_processor.prompt="an old vintage car" data.batch_size=16 trainer.visualize_progress=true

ProlificDreamer

python $FILE --config configs/prolificdreamer.yaml --train system.prompt_processor.prompt="a detailed Victorian era house" data.batch_size=[8,2] trainer.visualize_progress=true

python $FILE --config configs/prolificdreamer-geometry.yaml --train system.prompt_processor.prompt="a detailed Victorian era house" data.batch_size=8 trainer.visualize_progress=true system.geometry_convert_from=path/to/coarse/stage/trial/dir/ckpts/last.ckpt

python $FILE --config configs/prolificdreamer-texture.yaml --train system.prompt_processor.prompt="a detailed Victorian era house"  data.batch_size=1 trainer.visualize_progress=true system.geometry_convert_from=path/to/geometry/stage/trial/dir/ckpts/last.ckpt

DreamGaussian

Since DreamGaussian is a separate standalone repo, we have copied it in the dreamgaussian/ subfolder with our dreampropeller implementation specifically tailored to this repo. Please find specific instructions to run DreamGaussian inside.

Zero 1-to-3

Our method can also be applied to Image-to-3D using score distillation. Luckily Zero 1-to-3 is implemented in the original threestudio code, so we can directly plug in our method.

Download pretrained Zero123XL weights into load/zero123:

cd load/zero123
wget https://zero123.cs.columbia.edu/assets/zero123-xl.ckpt

To run,

python $FILE --config configs/zero123.yaml --train data.image_path=./load/images/grootplant_rgba.png data.random_camera.batch_size=[16,16,10] trainer.visualize_progress=true

Reference

@article{zhou2023dreampropeller,
  title={DreamPropeller: Supercharge Text-to-3D Generation with Parallel Sampling},
  author={Zhou, Linqi and Shih, Andy and Meng, Chenlin and Ermon, Stefano},
  journal={arXiv preprint arXiv:2311.17082},
  year={2023}
}