3D-Fixer: Coarse-to-Fine In-place Completion for 3D Scenes from a Single Image

Project Page | Paper | GitHub

3D-Fixer introduces a novel In-Place Completion paradigm to create high-fidelity 3D scenes from a single image. It extends 3D object generative priors to generate complete 3D assets conditioned on partially visible point clouds, using fragmented geometry as a spatial anchor to preserve layout fidelity without the need for time-consuming pose optimization.

Sample Usage

The model can be loaded using the following code snippet from the official repository:

The root config.yaml lists the released components, alternative checkpoint formats, file sizes and SHA256 digests. pipeline.json remains the runtime configuration. Download the release catalog with the Safetensors checkpoints and load the local directory:

from huggingface_hub import snapshot_download
from threeDFixer.pipelines import ThreeDFixerPipeline

model_dir = snapshot_download(
    "HorizonRobotics/3D-Fixer",
    allow_patterns=["config.yaml", "pipeline.json", "ckpts/*.json", "ckpts/*.safetensors"],
    local_dir="3D-Fixer",
)
pipeline = ThreeDFixerPipeline.from_pretrained(model_dir)

The .pt files are alternative checkpoint files and are not needed for this example. The pipeline also loads the external models named in pipeline.json; follow the official repository's setup instructions for those dependencies.

Dataset

The model was trained on ARSG-110K, a large-scale scene-level dataset comprising over 110K diverse scenes and 3M annotated images with high-fidelity 3D ground truth.

Citation

@inproceedings{yin2026tdfixer,
  title={3D-Fixer: Coarse-to-Fine In-place Completion for 3D Scenes from a Single Image},
  author={Yin, Ze-Xin and Liu, Liu and Wang, Xinjie and Sui, Wei and Su, Zhizhong and Yang, Jian and Xie, jin},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  year={2026}
}