TL;DR: GENA3D bridges 2D amodal completion and 3D generative modeling to achieve amodal 3D objects generation from sparse and paritial-occluded observations.

Official implementation for paper 'GENA3D: Generative Amodal 3D Modeling by Bridging 2D Priors and 3D Coherence'.

Intergrating 2D amodal completion prior and 3D generative modeling ability in the latent 3D space to achieve amodal 3D objects generation from sparse and partial-occluded views, under various scenarios, including single object-level, in-the-wild and in-the-scene.

https://github.com/user-attachments/assets/6f4b36e1-c50d-436a-9241-bd0e700c809e

Updates and TODOs

• ✔️ 12/04/2025: Initialize the project page.
• 🎊 06/18/2026: GENA3D is accepted to ECCV 2026.
• ✔️ 06/27/2026: Inference code and pretrained weights released!
• 🔲 TODO: Training code and training set construction.

Setup

Thanks Amodal3R for providing the environment setup and follow exactly as their instruction in this work.

Create a new conda environment named gena3d and install the dependencies:

. ./setup.sh --new-env --basic --xformers --flash-attn --diffoctreerast --spconv --mipgaussian --kaolin --nvdiffrast

The detailed usage of setup.sh can be found by running . ./setup.sh --help.

Usage: setup.sh [OPTIONS]
Options:
    -h, --help              Display this help message
    --new-env               Create a new conda environment
    --basic                 Install basic dependencies
    --train                 Install training dependencies
    --xformers              Install xformers
    --flash-attn            Install flash-attn
    --diffoctreerast        Install diffoctreerast
    --vox2seq               Install vox2seq
    --spconv                Install spconv
    --mipgaussian           Install mip-splatting
    --kaolin                Install kaolin
    --nvdiffrast            Install nvdiffrast
    --demo                  Install all dependencies for demo

Pretrained models

We have provided our pretrained weights on HuggingFace.

Inference

We have prepared demos under ./examples folder, supporting both single-view and sparse-view demos, you can run:

python inference_gena3d.py                      # no arg -> uses examples/demo1
python inference_gena3d.py --input examples/demo4   # single view demo
python inference_gena3d.py --input examples/demo3   # sparse view demo

If you want to try out on your own exampels:

First of all, we need to get the 2D completion methods using exisitng 2D frontends. We recommend OAAD for the same setting as in our paper. Also you can try out any other 2D amodal completion frontends. Apply it to all views of your case and compose them just as the structure shown in the demos.

Please check out the file formats needed for inference in demos. Basically we need amodal_completion.png (object after 2D amodal completion), occ_mask0.png (occlusion mask for SLAT generation), sd_img_cut.png (original occluded object image), visibility_instance_mask.png (visibilty mask, can be generated through sd_img_cut).

Training & Data Preparation

Coming soon...

Method

GENA3D bridges the 2D amodal completion with 3D generation using deliberaely designed View-Wise Cross Attention and Stereo-Conditioned Cross Attention in the Sparse Structure Generation Stage, with synthesized sparse-view 3D consistent occlusions as training data.

A detailed illustration of our proposed ViewWise Cross Attention and Stereo-Conditioned Cross Attention modules.

Results

Results on GSO object-level synthetic dataset.

Results on in-the-wild real-world captures.

Citation

Here is the bibtex reference. If you find our work interesting or useful, please give us a :star: or cite our paper!

@misc{zhou2026gena3dgenerativeamodal3d,
      title={GENA3D: Generative Amodal 3D Modeling by Bridging 2D Priors and 3D Coherence}, 
      author={Junwei Zhou and Yu-Wing Tai},
      year={2026},
      eprint={2511.21945},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.21945}, 
}