Home

Awesome

<br> <div align="center"> <img src="assets/logo.jpg" width="400px"></img> <h2>Training-free Acceleration <br> for 3D Generation 🏎️💨<h2> <p align="center"> <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-4caf50.svg" alt="License"></a> <a href="https://arxiv.org/abs/2404.06091" target="_blank"><img src="https://img.shields.io/badge/arXiv-2404.06091-b31b1b.svg" alt="arXiv"> <a href="https://adamdad.github.io/hash3D/" target="_blank"><img src="https://img.shields.io/badge/Project-Website-orange" alt="arXiv"></a> </p> </div>

Introduction

This repository contains the official implementation for our paper

Hash3D: Training-free Acceleration for 3D Generation

🥯[Project Page] 📝[Paper] </>[code]

Xingyi Yang, Xinchao Wang

National University of Singapore

pipeline

We present Hash3D, a universal solution to acclerate score distillation samplin (SDS) based 3D generation. By effectively hashing and reusing these feature maps across neighboring timesteps and camera angles, Hash3D substantially prevents redundant calculations, thus accelerating the diffusion model's inference in 3D generation tasks.

What we offer:

Results Visualizations

<section class="section"> <div class="container is-max-desktop"> <div class="columns is-centered"> <div class="column is-full-width"> <h3 class="title is-3">Image-to-3D Results</h3> <div class="content has-text-justified"> <table> <tr> <th>Input Image</th> <th>Zero-1-to-3</th> <th>Hash3D + Zero-1-to-3 $${\color{red} \text{(Speed X4.0)}}$$ </th> </tr> <tr> <td>

baby_phoenix_on_ice (1)

</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/797d78f0-d2d7-43a3-94af-bf57c9c5ef70

</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/c02701f1-fd92-4601-8569-18c7c17cde97

</td> </tr> <tr> <td>

grootplant_rgba (1)

</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/a41ba688-40bf-4d95-95de-37b669a90887

</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/86d9e46d-0554-4a87-9960-ce3a9f83bdd7

</td> </tr> </table> </div> </div> </div> </div> <section class="section"> <div class="container is-max-desktop"> <div class="columns is-centered"> <div class="column is-full-width"> <h3 class="title is-3">Text-to-3D Results</h3> <div class="content has-text-justified"> <table> <tr> <th>Prompt</th> <th>Gaussian-Dreamer</th> <th>Hash3D + Gaussian-Dreamer $${\color{red}\text{(Speed X1.5)}}$$ </th> </tr> <tr> <td style="text-align: center; vertical-align: middle; font-size:18px; font-family:Comic sans MS">A bear dressed as a lumberjack</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/80a4658f-7233-49aa-a357-ff296396185b

</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/3882341f-c5f1-4f4f-8f24-d1c080ecdb2f

</td> </tr> <tr> <td style="text-align: center; vertical-align: middle; font-size:18px; font-family:Comic sans MS">A train engine made out of clay</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/1111d8ba-aae5-4117-9340-5d950702e49b

</td> <td>

https://github.com/Adamdad/hash3D/assets/26020510/06b7bbf3-0edb-4d2f-a2f2-c11bab5c7b64

</td> </tr> </table> </div> </div> </div> </div>

Project Structure

The repository is organized into three main directories, each catering to a different repo that Hash3D can be applied on:

  1. threesdtudio-hash3d: Contains the implementation of Hash3D tailored for use with the threestudio.
  2. dreamgaussian-hash3d: Focuses on integrating Hash3D with the DreamGaussian for image-to-3D generation.
  3. gaussian-dreamer-hash3d: Dedicated to applying Hash3D to GaussianDreamer for faster text-to-3D tasks.

What we add?

The core implementation is in the guidance_loss for each SDS loss computation. We

See hash3D/threestudio-hash3d/threestudio/models/guidance/zero123_unified_guidance_cache.py for example. The code for the hash table implementation is in hash3D/threestudio-hash3d/threestudio/utils/hash_table.py.

Getting Started

Installation

Navigate to each of the specific directories for environment-specific installation instructions.

Usage

Refer to the README within each directory for detailed usage instructions tailored to each environment.

For example, to run Zero123+SDS with hash3D

cd threestudio-hash3d
python launch.py --config configs/stable-zero123_hash3d.yaml --train --gpu 0 data.image_path=https://adamdad.github.io/hash3D/load/images/dog1_rgba.png

Evaliation

  1. Image-to-3D: GSO dataset GT meshes and renderings can be found online. With the rendering of the reconstructed 3D objects at pred_dir and the gt rendering at gt_dir, run
python eval_nvs.py --gt $gt_dir --pr $pred_dir 
  1. Text-to-3D: Run all the prompts in assets/prompt.txt. And compute the CLIP score between text and rendered image as
python eval_clip_sim.py "$gt_prompt" $pred_dir --mode text

Acknowledgement

We borrow part of the code from DeepCache for feature extraction from diffusion models. We also thanks the implementation from threestudio, DreamGaussian, Gaussian-Dreamer, and the valuable disscussion with @FlorinShum and @Horseee.

Citation

@misc{yang2024hash3d,
      title={Hash3D: Training-free Acceleration for 3D Generation}, 
      author={Xingyi Yang and Xinchao Wang},
      year={2024},
      eprint={2404.06091},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}