ParthaEth / GIF

GIF is a photorealistic generative face model with explicit 3D geometric and photometric control.
https://gif.is.tue.mpg.de/
MIT License
404 stars 63 forks source link
conditonal-gan gan gans generative-adversarial-network generative-interpretable-faces generative-models latent-variable-models photorealism photorealistic-faces

GIF: Generative Interpretable Faces

This is the official implementation for the paper GIF: Generative Interpretable Faces. GIF is a photorealistic generative face model with explicit control over 3D geometry (parametrized like FLAME), appearance, and lighting.

Important links

Watch a brief presentation

Watch a presentation

Citation

If you find our work useful in your project please cite us as

@inproceedings{GIF2020,
    title = {{GIF}: Generative Interpretable Faces},
    author = {Ghosh, Partha and Gupta, Pravir Singh and Uziel, Roy and Ranjan, Anurag and Black, Michael J. and Bolkart, Timo},
    booktitle = {International Conference on 3D Vision (3DV)},
    year = {2020},
    url = {http://gif.is.tue.mpg.de/}
}

Installation

First thing first

Before Running any program you will need to download a few resource files and create a suitable placeholder for the training artifacts to be stored

  1. you can use this link to download input files necessary to train GIF from scratch - http://files.is.tuebingen.mpg.de/gif/input_files.zip
  2. you can use this link to download checkpoints and samples generated from pre-trained GIF models and its ablated versions - http://files.is.tuebingen.mpg.de/gif/output_files.zip
  3. Now create a directory called GIF_resources and unzip the input zip or checpoint zip or both in this directory
  4. When you train or fine tune a model the output directory checkpoint and sample directory will be populated. Rmember that the model atifacts can easily become a few 10s of terabytes
  5. The main resource directory should be named GIF_resources and it should have input_files and output_fiels as sub-directories
  6. Now you need to provide the path to this directory in the constants.py script and make changes if necessary if you wish to change names of the subdirectories
  7. Edit the line resources_root = '/path/to/the/unzipped/location/of/GIF_resources'
  8. Modify any other paths as you need
  9. Download the FLAME 2020 model and the FLAME texture space from here - https://flame.is.tue.mpg.de/ (you need to sign up and agree to the license for access)
  10. Please make sure to dowload 2020 version. After signing in you sould be able to download FLAME 2020
  11. Download the FLAME_texture_data, Unzip it and place the texture_data_256.npy file in the flame resources directory.
  12. Please place the generic_model.pkl file in GIF_resources/input_files/flame_resource
  13. In this directory you will need to place the generic_model.pkl, head_template_mesh.obj, and FLAME_texture.npz in addition to the already provided files in the zip you just downloaded from the link given above. You can find these files from the official flame website. Link given in point 9.

Preparing training data

To train GIF you will need to prepare two lmdb datasets

  1. An LMDB datset containing FFHQ images in different scales
    1. To prepare this cd prepare_lmdb
    2. run python prepare_ffhq_multiscale_dataset.py --n_worker N_WORKER DATASET_PATH
    3. Here DATASET_PATH is the parth to the directory that contains the FFHQ images
    4. Place the created lmdb file in the GIF_resources/input_files/FFHQ directory, alongside ffhq_fid_stats
  2. An LMDB dataset containing renderings of the FLAME model
    1. To run GIF you will need the rendered texture and normal images of the FLAME mesh for FFHQ images. This is already provided as deca_rendered_with_public_texture.lmdb with the input_file zip. It is located in GIF_resources_to_upload/input_files/DECA_inferred
    2. To create this on your own simply run python create_deca_rendered_lmdb.py

Training

To resume training from a checkpoint run python train.py --run_id <runid> --ckpt /path/to/saved.mdl/file/<runid>/model_checkpoint_name.model

Note here that you point to the .model file not the npz one.

To start training from scratch run python train.py --run_id <runid>

Note that the training code will take all available GPUs in the system and perform data parallelization. You can set visible GPUs by etting the CUDA_VISIBLE_DEVICES environment variable. Run CUDA_VISIBLE_DEVICES=0,1 python train.py --run_id <runid> to run on GPU 0 and 1

To run random face generation follow the following steps

  1. Clone this repo
  2. Download the pretrained model. Please note that you have to download the model with correct run_id
  3. activate your virtual environment
  4. cd plots
  5. python generate_random_samples.py
  6. Remember to uncomment the appropriate run_id

To generate Figure 3

  1. cd plots
  2. python role_of_different_parameters.py it will generate batch_size number of directories in f'{cnst.output_root}sample/' named gen_iamges<batch_idx>. Each of these directory contain a column of images as shown in figure 3 in the paper.

Amazon mechanical turk (AMT) evaluations:

Disclaimer: This section can be outdated and/or have changed since the time of writing the document. It is neither intended to advertise nor recommend any particular 3rd party product. The inclusion of this guide is solely for quick reference purposes and is provided without any liability.

Running the naive vector conditioning model

Acknowledgements

GIF uses DECA to get FLAME geometry, appearance, and lighting parameters for the FFHQ training data. We thank H. Feng for prepraring the training data, Y. Feng and S. Sanyal for support with the rendering and projection pipeline, and C. Köhler, A. Chandrasekaran, M. Keller, M. Landry, C. Huang, A. Osman and D. Tzionas for fruitful discussions, advice and proofreading. We specially thank Taylor McConnell for voicing over our video. The work was partially supported by the International Max Planck Research School for Intelligent Systems (IMPRS-IS) and by Amazon Web Services.