[CVPR 2026] LaS-Comp: Zero-shot 3D Completion with Latent–Spatial Consistency

Official implementation of "LaS-Comp: Zero-shot 3D Completion with Latent–Spatial Consistency"

Weilong Yan, Haipeng Li, Hao Xu, Nianjin Ye, Yihao Ai, Shuaicheng Liu, Jingyu Hu

• [2026-04] 🎉 LaS-Comp is selected to do a poster presentation in CHINA3DV 2026!
• [2026-03] 🎉 Our proposed benchmark Omni-Comp3D is now available at Hugging Face!
• [2026-02] 📄 Our work is accepted by CVPR 2026, and paper is available on arXiv.

This project is tested under the following environment:

• Python: 3.10
• CUDA: 12.1
• PyTorch: 2.4.0 (compiled with CUDA 12.1)
• torchvision: 0.19.0

⚠️ Important: Please ensure your system CUDA version is 12.1.
Mismatched CUDA versions (e.g., 11.x or 12.3+) may cause errors with spconv, flash-attn, or rendering modules.

conda env create -f environment.yml
conda activate lascomp

pip install -r requirements.txt

python -c "import torch; print(torch.cuda.is_available())"

This project relies on several CUDA-dependent libraries:

• spconv-cu121
• flash-attn
• nvdiffrast
• diff-gaussian-rasterization

Some dependencies (e.g., xformers, kaolin) may require manual installation depending on your system.

• Release Omni-Comp3D dataset.
• Release code for TRELLIS.
• Release code for Direct3D-S2.

Following the official TRELLIS release, please download the pretrained checkpoints for image-large and text-xlarge, and place them under the ckpt/ directory. The directory structure should look like:

ckpt/
├── image-large/
├── text-xlarge/
└── clip/

The Omni-Comp3D benchmark is hosted on Hugging Face. It includes two parts:

• Omni-Comp3D: our proposed benchmark for comprehensive completion evaluation.
• samples: the evaluation samples following SDS-Complete, GenPC, ComPC.

You can access them here: 👉 Omni-Comp3D

Please place both Omni-Comp3D and samples in their appropriate locations within the project directory.

For custom partial inputs, you can run the project in either text-conditioned or image-conditioned mode.

python run_lascomp_text_condition_single.py \
  --partial-path path/to/your/partial-shape \
  --prompt "Your prompt" \
  --dataset custom \
  --yz-flip

python run_lascomp_image_condition_single.py \
  --partial-path path/to/your/partial-shape \
  --image-path path/to/your/image \
  --dataset custom \
  --yz-flip

• --partial-path specifies the input partial 3D shape.
• --prompt is used for the text-conditioned model.
• --image-path is used for the image-conditioned model.
• --dataset custom indicates that the input comes from your own custom data.
• --yz-flip is used for samples whose vertical axis is y. If your sample uses z as the default vertical axis, please use --no-yz-flip instead.
• You can adjust the hyperparameters to get better completion results.

To evaluate the model on the provided benchmarks, you can use the following scripts for different datasets.

python run_lascomp_text_condition_omnicomp.py

python run_lascomp_image_condition.py

or

python run_lascomp_text_condition.py

We sincerely thank TRELLIS, Direct3D-S2, ComPC, FlowChef, FlowDPS, and VoxHammer for their inspirational help to our work.

If you find our work or dataset helpful for your research, please consider citing:

@InProceedings{Yan_2026_CVPR,
    author    = {Yan, Weilong and Li, Haipeng and Xu, Hao and Ye, Nianjin and Ai, Yihao and Liu, Shuaicheng and Hu, Jingyu},
    title     = {LaS-Comp: Zero-shot 3D Completion with Latent-Spatial Consistency},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {7588-7599}
}