🛶SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering

This repository contains official code implementation of "SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering". We provide Diffusers-based demo for simple testing. The results may differ from those reported in the paper.

Setup

1. Conda setup - CUDA 12.1 with A100 80GB GPU

   conda create -n steerx python=3.10 -y
   conda activate steerx
   pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu121
   pip install --extra-index-url https://miropsota.github.io/torch_packages_builder pytorch3d==0.7.8+pt2.4.1cu121
   pip install -r requirements.txt

2. Download checkpoints - we collect all checkpoints.

gdown --folder https://drive.google.com/drive/folders/1YrsZq54YMXKtNLwxqOC7yAx_nYlyymPo?usp=drive_link

3. (Optional for faster runtime) Compile the cuda kernels for RoPE.

cd steerx_diffusers/geometry_steering/croco/models/curope/
python setup.py build_ext --inplace
cd ../../../../../

4. Requirements for num_frames and resolution. As default, we set num_frames = 49 for 4D Generation.

• Mochi : 6k + 1, min(H, W) = 480
• Hunyuan : 4k + 1, min(H, W) = 480
• CogVideoX : 16k + 1, min(H, W) = 768
• DimensionX: 49, 720 x 480
• SplatFlow: 8, 256 x 256

Run code (--img_path is only required for image-to-3D/4D)

python eval_steerx.py --model {generative model name} --num_particles {num_particles}
python demo_steerx.py --model {generative model name} --num_particles {num_particles} --prompt {your_prompt} --img_path {img_path}

Visualization

3D Gaussian Splatting.

Refer to superspl.at/editor

4D Pointmaps. You may adjust vis setup for better results.

python viser/visualizer_monst3r.py --data {your data path}

Acknowledgments

Thanks for wonderful works to build this repo.

• Video generative models in Mochi, HunyuanVideo, CogVideoX, DimensionX, and SplatFlow
• Scene reconstruction models in MV-DUSt3R+, MonST3R, and SplatFlow.
• Reward functions are based on MEt3R
• Steering is based on FKS.

Citation

If you find this repository helpful for your project, please consider citing our work. :)

@article{park2025steerx,
  title={SteerX: Creating Any Camera-Free 3D and 4D Scenes with Geometric Steering},
  author={Park, Byeongjun and Go, Hyojun and Nam, Hyelin and Kim, Byung-Hoon and Chung, Hyungjin and Kim, Changick},
  journal={arXiv preprint arXiv:2503.12024},
  year={2025}
}