Visual Object Networks

We present Visual Object Networks (VON), an end-to-end adversarial learning framework that jointly models 3D shapes and 2D images. Our model can synthesize a 3D shape, its intermediate 2.5D depth representation, and a final 2D image all at once. The VON not only generates images that are more realistic than recent 2D GANs, but also enables several 3D operations.


Visual Object Networks: Image Generation with Disentangled 3D Representation.
Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Joshua B. Tenenbaum, William T. Freeman.
MIT CSAIL and Google Research.
In NeurIPS 2018.

Example results

(a) Typical examples produced by a recent GAN model [Gulrajani et al., 2017].

(b) Our model produces 3 outputs: a 3D shape, its 2.5D projection given a viewpoint, and a final image with realistic texture.

(c) Given this disentangled 3D representation, our method allows several 3D applications including editing viewpoint, shape, or texture independently.


More Samples

Below we show more samples from DCGAN [Radford et al., 2016], LSGAN [Mao et al., 2017], WGAN-GP [Gulrajani et al., 2017], and our VON. For our method, we show both 3D shapes and 2D images. The learned 3D prior helps our model produce better samples.


3D Object Manipulations

Our Visual Object Networks (VON) allow several 3D applications such as (left) changing the viewpoint, texture, or shape independently, and (right) interpolating between two objects in shape space, texture space, or both.


Texture Transfer across Objects and Viewpoints

VON can transfer the texture of a real image to different shapes and viewpoints



  • Linux (only tested on Ubuntu 16.04)
  • Python3 (only tested with python 3.6)
  • Anaconda3

Getting Started


  • Clone this repo:
git clone -b master --single-branch
cd VON
  • Install PyTorch 0.4.1+ and torchvision from and other dependencies (e.g., visdom and dominate). You can install all the dependencies by the following:
conda create --name von --file pkg_specs.txt
source activate von
  • (Optional) Install blender for visualizing generated 3D shapes. After installation, please add blender to your PATH environment variable.

Generate 3D shapes, 2.5D sketches, and images

  • Download our pretrained models:
bash ./datasets/

-generate results with the model

bash ./scripts/ car

The test results will be saved to a html file here: ./results//val/index.html.

Model Training

  • To train a model, download the training dataset(distance functions and images).
bash ./datasets/
  • Training 3D generative model:
bash ./scripts/
  • Training 2D image generation using ShapeNet objects:
bash ./scripts/
  • Train 2D image generation models using trained 3D generator:
bash ./scripts/
  • Jointly finetune 3D and 2D generative models:
bash ./scripts/
  • To view training results and loss plots, go to http://localhost:8097 in a web browser. To see more intermediate results, check out ./checkpoints/*/web/index.html


If you find this useful for your research, please cite the following paper.

  title={Visual Object Networks: Image Generation with Disentangled 3{D} Representations},
  author={Jun-Yan Zhu and Zhoutong Zhang and Chengkai Zhang and Jiajun Wu and Antonio Torralba and Joshua B. Tenenbaum and William T. Freeman},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},


This work is supported by NSF #1231216, NSF #1524817, ONR MURI N00014-16-1-2007, Toyota Research Institute, Shell, and Facebook. We thank Xiuming Zhang, Richard Zhang, David Bau, and Zhuang Liu for valuable discussions. This code borrows from the CycleGAN & pix2pix repo.