PHysGen3D: Crafting a Miniature Interactive
World from a Single Image

CVPR, 2025
Boyuan Chen · Hanxiao Jiang · Shaowei Liu · Saurabh Gupta · Yunzhu Li · Hao Zhao · Shenlong Wang

This repository contains the implementation for the paper PHysGen3D: Crafting a Miniature Interactive World from a Single Image, CVPR 2025. In this paper, we present a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene.

Overview

Structure

The folders are exsisting wheels used in the projects. "engine" folder contains the core of taichi-elements.

Run perception.py to run the perception part.

Run ball_sim.py mpm_sim.py to run several demos of mpm method.

Installation

conda create -y -n phys python=3.10
conda activate phys
git clone --recurse-submodules git@github.com:by-luckk/PhysGen3D.git
cd PhysGen3D
bash env_install/env_install.sh
bash env_install/download_pretrained.sh

## Usage

The examples below are provided for the demo images in `data/img/`. The `teddy.jpg` can be substituted with any other images. `${name}` is the name of the image.

### Run the perception part

```bash
python perception.py --input_image data/img/teddy.jpg --text_prompt teddy

• The text prompt discribes the object you want to move. It's in format of a single word or multiple words seperated by . like cat.dog.

• Outputs are saved in outputs/${name} as follows:

  ${name}/
    ├── depth # Depth point cloud
    ├── images # Multiview object images
    ├── inpaint # Background inpainting
    ├── mask # Object masks
    ├── meshes # Mesh reconstruction
    ├── object # Object registration results
    ├── grounded_sam_output.jpg
    ├── raw_image.jpg
    └── transform.json # Geometries
  

Run the simulation part

python simulation.py --config data/sim/teddy.yaml

• You can manually set all the physical parameters or get them automatically using GPT-4o.
• Velocities is the initial velocity of object(s), in 1D or 2D array: [Vx, Vy, Vz] or [[Vx1, Vy1, Vz1], [Vx2, Vy2, Vz2]].
• The outputs are saved in sim_result/sim_result_${time} folder.

Run the rendering part

python rendering.py \
-i ./sim_result/sim_result_${time} \
--path outputs/teddy \
--env data/hdr/teddy.exr \
-b 0 \
-e 100 \
-f \
-s 1 \
-o render_result/1 \
-M 460 \
-p 20 \
--shutter-time 0.0

• In run_mitsuba.sh, put your simulation results folder sim_result/sim_result_${time} after -i.
• It runs the teddy bear demo by default. For other demos, put perception result outputs/${name} after --path and env light file data/hdr/teddy.exr after --env.
• The outputs and the final video are saved in render_result folder.

Prepare your own image