Controllable and Compositional Generation with Latent-Space Energy-Based Models
Official PyTorch implementation of the NeurIPS 2021 paper:
Controllable and Compositional Generation with Latent-Space Energy-Based Models
Weili Nie, Arash Vahdat, Anima Anandkumar
https://nvlabs.github.io/LACE
Abstract: Controllable generation is one of the key requirements for successful adoption of deep generative models in real-world applications, but it still remains as a great challenge. In particular, the compositional ability to generate novel concept combinations is out of reach for most current models. In this work, we use energy-based models (EBMs) to handle compositional generation over a set of attributes. To make them scalable to high-resolution image generation, we introduce an EBM in the latent space of a pre-trained generative model such as StyleGAN. We propose a novel EBM formulation representing the joint distribution of data and attributes together, and we show how sampling from it is formulated as solving an ordinary differential equation (ODE). Given a pre-trained generator, all we need for controllable generation is to train an attribute classifier. Sampling with ODEs is done efficiently in the latent space and is robust to hyperparameters. Thus, our method is simple, fast to train, and efficient to sample. Experimental results show that our method outperforms the state-of-the-art in both conditional sampling and sequential editing. In compositional generation, our method excels at zero-shot generation of unseen attribute combinations. Also, by composing energy functions with logical operators, this work is the first to achieve such compositionality in generating photo-realistic images of resolution 1024x1024.
Requirements
- Linux and Windows are supported, but we recommend Linux for performance and compatibility reasons.
- 1 high-end NVIDIA GPU with at least 24 GB of memory. We have done all testing and development using a single NVIDIA V100 GPU with memory size 32 GB.
- 64-bit Python 3.8.
- CUDA=10.0 and docker must be installed first.
- Installation of the required library dependencies with Docker:
docker build -f lace-cuda-10p0.Dockerfile --tag=lace-cuda-10-0:0.0.1 . docker run -it -d --gpus 0 --name lace --shm-size 8G -v $(pwd):/workspace -p 5001:6006 lace-cuda-10-0:0.0.1 docker exec -it lace bash
Experiments on CIFAR-10
The CIFAR10
folder contains the codebase to get the main results on the CIFAR-10 dataset, where the scripts
folder contains the necessary bash scripts to run the code.
Data preparation
Before running the code, you have to download the data (i.e., the latent code and label pairs) from here and unzip it to the CIFAR10
folder. Or you can go to the folder CIFAR10/prepare_data
and follow the instructions to generate the data.
Training
To train the latent classifier, you can run:
bash scripts/run_clf.sh
In the script run_clf.sh
, the variable x
can be specified to w
or z
, representing that the latent classifier is trained in the w-space or z-space of StyleGAN, respectively.
Sampling
To get the conditional sampling results with the ODE or Langevin dynamics (LD) sampler, you can run:
# ODE
bash scripts/run_cond_ode_sample.sh
# LD
bash scripts/run_cond_ld_sample.sh
By default, we set x
to w
, meaning we use the w-space classifier, because we find our method works the best in w-space. You can change the value of x
to z
or i
to use the classifier in z-space or pixel space, for a comparison.
To compute the conditional accuracy (ACC) and FID scores in conditional sampling with the ODE or LD sampler, you can run:
# ODE
bash scripts/run_cond_ode_score.sh
# LD
bash scripts/run_cond_ld_score.sh
Note that:
-
For the ACC evaluation, you need a pre-trained image classifier, which can be downloaded as instructed here;
-
For the FID evaluation, you need to have the FID reference statistics computed beforehand. You can go to the folder
CIFAR10/prepare_data
and follow the instructions to compute the FID reference statistics with real images sampled from CIFAR-10.
Experiments on FFHQ
The FFHQ
folder contains the codebase for getting the main results on the FFHQ dataset, where the scripts
folder contains the necessary bash scripts to run the code.
Data preparation
Before running the code, you have to download the data (i.e., 10k pairs of latent variables and labels) from here (originally from StyleFlow) and unzip it to the FFHQ
folder.
Training
To train the latent classifier, you can run:
bash scripts/run_clf.sh
Note that each att_name
(i.e., glasses) in run_clf.sh
corresponds to a separate attribute classifier.
Sampling
First, you have to get the pre-trained StyleGAN2 (config-f) by following the instructions in Convert StyleGAN2 weight from official checkpoints.
Conditional sampling
To get the conditional sampling results with the ODE or LD sampler, you can run:
# ODE
bash scripts/run_cond_ode_sample.sh
# LD
bash scripts/run_cond_ld_sample.sh
To compute the conditional accuracy (ACC) and FID scores in conditional sampling with the ODE or LD sampler, you can run:
# ODE
bash scripts/run_cond_ode_score.sh
# LD
bash scripts/run_cond_ld_score.sh
Note that:
-
For the ACC evaluation, you need to train an FFHQ image classifier, as instructed here;
-
For the FID evaluation, you need to have the FID reference statistics computed beforehand. You can go to the folder
FFHQ/prepare_models_data
and follow the instructions to compute the FID reference statistics with the StyleGAN generated FFHQ images.
Sequential editing
To get the qualitative and quantitative results of sequential editing, you can run:
# User-specified sampling
bash scripts/run_seq_edit_sample.sh
# ACC and FID
bash scripts/run_seq_edit_score.sh
Note that:
-
Similarly, you first need to train an FFHQ image classifier and get the FID reference statics to compute ACC and FID score by following the instructions, respectively.
-
To get the face identity preservation (ID) score, you first need to download the pre-trained ArcFace network, which is publicly available here, to the folder
FFHQ/pretrained/metrics
.
Compositional Generation
To get the results of zero-shot generation on novel attribute combinations, you can run:
bash scripts/run_zero_shot.sh
To get the results of compositions of energy functions with logical operators, we run:
bash scripts/run_combine_energy.sh
Experiments on MetFaces
The MetFaces
folder contains the codebase for getting the main results on the MetFaces dataset, where the scripts
folder contains the necessary bash scripts to run the code.
Data preparation
Before running the code, you have to download the data (i.e., 10k pairs of latent variables and labels) from here and unzip it to the MetFaces
folder. Or you can go to the folder MetFaces/prepare_data
and follow the instructions to generate the data.
Training
To train the latent classifier, you can run:
bash scripts/run_clf.sh
Note that each att_name
(i.e., yaw) in run_clf.sh
corresponds to a separate attribute classifier.
Sampling
To get the conditional sampling and sequential editing results, you can run:
# conditional sampling
bash scripts/run_cond_sample.sh
# sequential editing
bash scripts/run_seq_edit_sample.sh
Experiments on AFHQ-Cats
The AFHQ
folder contains the codebase for getting the main results on the AFHQ-Cats dataset, where the scripts
folder contains the necessary bash scripts to run the code.
Data preparation
Before running the code, you have to download the data (i.e., 10k pairs of latent variables and labels) from here and unzip it to the AFHQ
folder. Or you can go to the folder AFHQ/prepare_data
and follow the instructions to generate the data.
Training
To train the latent classifier, you can run:
bash scripts/run_clf.sh
Note that each att_name
(i.e., breeds) in run_clf.sh
corresponds to a separate attribute classifier.
Sampling
To get the conditional sampling and sequential editing results, you can run:
# conditional sampling
bash scripts/run_cond_sample.sh
# sequential editing
bash scripts/run_seq_edit_sample.sh
License
Please check the LICENSE file. This work may be used non-commercially, meaning for research or evaluation purposes only. For business inquiries, please contact [email protected].
Citation
Please cite our paper, if you happen to use this codebase:
@inproceedings{nie2021controllable,
title={Controllable and compositional generation with latent-space energy-based models},
author={Nie, Weili and Vahdat, Arash and Anandkumar, Anima},
booktitle={Neural Information Processing Systems (NeurIPS)},
year={2021}
}