viewset-diffusion

Implementation of `Viewset Diffusion: (0-)Image-Conditioned 3D Generative Models from 2D Data'.

This is the first release of the code, this README will get more comprehensive with time. Do report bugs if you spot them and contributions are always welcome.

Requirements

Install requirements and activate conda environment.

conda env create -f environment.yml
conda activate viewset_diffusion

To train Viewset Diffusion models you will need 1x24GB GPU for training on the Minens dataset and 2x48GB GPUs for training on CO3D and ShapeNet Cars.

Data

Minens dataset

Download files from skin_names to data/' so thatdata/' contains two folders: data/meta and `data/meta2'.
Run bash ./download.sh to download skins from https://planetminecraft.com and https://minecraftskins.com. It will scrape a list of skin names in data/meta* and then a list of skin PNGs in data/skins* and save them to data/skins and data/skins2, respectively.
Download files from benchmark_minens into data/benchmark_mine. Run python render_minecraft_dataset.py. It will rasterize images of articulated, textured Minecraft characters and save the dataset as a .npy file.
Set the value of data.data_path in config/dataset/minens.yaml to the absolute path of the directory where minens was rendered.

Files data/benchmark_mine/cameras_* and data/benchmark_mine/poses_* hold camera poses and character poses in each dataset, respectively. data/benchmark_mine/training_skins.txt and data/benchmark_mine/testing_skins.txt hold skin names of each character. data/benchmark_mine/testing_bkgds.npy holds the background colours in each file. data/benchmark_mine/random_order_of_sample_idxs.npy and data/benchmark_mine/ambiguous_pose.npy hold file indices used for evaluation.

CO3D

Download CO3D following instructions from official website. Put the path of the download directory in data_manager/co3d.py as CO3D_DATASET_ROOT. Set the value of data.data_path in config/dataset/co3d_hydrant.yaml to the absolute path of where you want to store a compressed, fast-to-access .npy version of the CO3D classes. The first time training is run, the data manager will form train, val and test dataset and store them in that path.

If you want to run quantitative evaluation and reproduce scores from the paper you also need to download the indexes of samples I used in evaluation. Download random_order_of_sample_idxs.npy for each class from the folders with models for CO3D Hydrants, CO3D Teddybears, CO3D Vases or CO3D Plants. Place the order of files for evaluation in the same folder as the other .npy files.

SRN-ShapeNet Cars

Download the ShapeNet-SRN Car dataset from PixelNeRF project page. Set the value of data.data_path in config/dataset/srn_cars.yaml to the absolute path of the SRN Dataset root. The data should have the structure SRN_ROOT/srn_cars/srn_cars/cars_{train,val,test}/{example_id}.

Training

Specifying arguments for training follows the hydra configuration syntax, for example:

python train.py +experiment=diffusion_128 +arch=up_3d_conv +dataset=srn_cars

Training options

+experiment: set to diffusion_128 for training on 128x128 resolution (CO3D, ShapeNet Cars) and diffusion_48 for training on 48x48 resolution (Minens).

optimization.n_iter: set to 104,000 for training on ShapeNet Cars, 200,000-300,000 for training on CO3D classes and 200,000 for training on Minens.

+arch: set to up_3d_conv for training Viewset Diffusion. triplanes can be used to train RenderDiffusion, more information on that coming soon.

+dataset+: set to co3d_{hydrant,plant,teddybear,vase} for CO3D classes, srn_cars for ShapeNet Cars and minens for Minens.

Pretrained models

ShapeNet and CO3D checkpoints are now released - enjoy experimentation! You can now download our models for all CO3D and ShapeNet classes.

Evaluation

Once a model is trained, 3D reconstruction can be run with:

python evaluation_qualitative.py [model_path] --dataset_name [dataset_name] --N_clean 1 --N_noisy 3 --cf_guidance 0.0 --seed 0

3D Generation can be run with:

python evaluation_qualitative.py [model_path] --dataset_name [dataset_name] --N_clean 0 --N_noisy 4 --cf_guidance -1.0 --seed 0

Quantitative evaluation of 3D reconstruction is run with

python evaluation_quantitative.py model_path=[model path]

BibTeX

If you find this code useful, please consider citing

@inproceedings{szymanowicz23viewset_diffusion,
      title={Viewset Diffusion: (0-)Image-Conditioned {3D} Generative Models from {2D} data},
      author={Stanislaw Szymanowicz and Christian Rupprecht and Andrea Vedaldi},
      year={2023},
      booktitle={ICCV},
}

szymanowiczs / viewset-diffusion

readme