TLDR: A unified framework to explain SDS and its variants, plus a new method that is fast & high-quality.

We provide our code for text-based NeRF optimization as an extension in Threestudio. To use it, please first install threestudio following the official instructions.

cp -r ./threestudio-sds-bridge ../threestudio/custom/
cd ../threestudio

In the threestudio repo...

python launch.py --config custom/threestudio-sds-bridge/configs/sds-bridge.yaml --train --gpu 0 system.prompt_processor.prompt="a pineapple"

Some options to play with for sds-bridge guidance are:

• system.guidance.stage_two_start_step The step at which to switch to the second stage.
• system.guidance.stage_two_weight The weight of the second stage.
• system.prompt_processor.src_modifier The prompt modfier that describes the current source distribution, e.g. "oversaturated, smooth, pixelated, cartoon, foggy, hazy, blurry, bad structure, noisy, malformed."
• system.prompt_processor.tgt_modifier The prompt modfier that describes the target distribution, e.g. " detailed, high resolution, high quality, sharp."

We offer a simpler installation than Threestudio with minimal dependencies if you just want to run experiments in 2D. This installation guide is adapted from Nerfstudio

You must have an NVIDIA video card with CUDA installed on the system. This project has been tested with version 11.8 of CUDA. You can find more information about installing CUDA here

This repository requires python >= 3.8. We recommend using conda to manage dependencies. Make sure to install Conda before proceeding.

conda create --name bridge -y python=3.8
conda activate bridge
pip install --upgrade pip

Install PyTorch with CUDA

For CUDA 11.8:

pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

Install other dependencies with pip:

cd 2D_experiments
pip install -r requirements.txt

In the 2D_experiments directory...

python generate.py

See generate.py for more options, including but not limited to:

• --mode Choose between SDS-like loss functions bridge (ours), SDS, NFSD, VSD
• --seed Random seed
• --lr Learning rate
• --cfg_scale Scale of classifier-free guidance computation

@article{mcallister2024rethinking,
    title={Rethinking Score Distillation as a Bridge Between Image Distributions},
    author={David McAllister and Songwei Ge and Jia-Bin Huang and David W. Jacobs and Alexei A. Efros and Aleksander Holynski and Angjoo Kanazawa},
    journal={arXiv preprint arXiv:2406.09417},
    year={2024}
  }