pix2latent: framework for inverting images into generative models

Framework for inverting images. Codebase used in:

Transforming and Projecting Images into Class-conditional Generative Networks
project page | paper
Minyoung Huh   Richard Zhang   Jun-Yan Zhu   Sylvain Paris   Aaron Hertzmann
MIT CSAIL   Adobe Research
ECCV 2020 (oral)

@inproceedings{huh2020ganprojection,
    title = {Transforming and Projecting Images to Class-conditional Generative Networks},
    author = {Minyoung Huh and Richard Zhang, Jun-Yan Zhu and Sylvain Paris and Aaron Hertzmann},
    booktitle = {ECCV},
    year = {2020}
}

NOTE [8/25/20] The codebase has been renamed from GAN-Transform-and-Project to pix2latent, and also refactored to make it easier to use and extend to any generative model beyond BigGAN. To access the original codebase refer to the legacy branch.

Example results

All results below are without fine-tuning.

BigGAN (z-space) - ImageNet (256x256)

StyleGAN2 (z-space) - LSUN Cars (384x512)

StyleGAN2 (z-space) - FFHQ (1024x1024)

Prerequisites

The code was developed on

• Ubuntu 18.04
• Python 3.7
• PyTorch 1.4.0

Getting Started

• Install PyTorch
  Install the correct PyTorch version for your machine

• Install the python dependencies
  Install the remaining dependencies via

  pip install -r requirements.txt
  

• Install pix2latent

  git clone https://github.com/minyoungg/pix2latent
  cd pix2latent
  pip install .
  

Examples

We provide several demo codes in ./examples/ for both BigGAN and StyleGAN2. Note that the codebase has been tuned and developed on BigGAN.

> cd examples
> python invert_biggan_adam.py --num_samples 4

Using the make_video flag will save the optimization trajectory as a video.

> python invert_biggan_adam.py --make_video --num_samples 4

(slow) To optimize with CMA-ES or BasinCMA, we use PyCMA. Note that the PyCMA version of CMA-ES has a predefined number of samples to jointly evaluate (18 for BigGAN) and (22 for StyleGAN2).

> python invert_biggan_cma.py 
> python invert_biggan_basincma.py 

(fast) Alternatively CMA-ES in Nevergrad provides sample parallelization so you can set your own number of samples. Although this runs faster, we have observed the performance to be slightly worse. (warning: performance depends on num_samples).

> python invert_biggan_nevergrad.py --ng_method CMA --num_samples 4
> python invert_biggan_hybrid_nevergrad.py --ng_method CMA --num_samples 4

Same applies to StyleGAN2. See ./examples/ for extensive list of examples.

Template pseudocode

import torch, torch.nn as nn
import pix2latent.VariableManger
from pix2latent.optimizer import GradientOptimizer

# load your favorite model
class Generator(nn.Module):
    ...
    
    def forward(self, z):
        ...
        return im

model = Generator() 

# define your loss objective .. or use the predefined loss functions in pix2latent.loss_functions
loss_fn = lambda out, target: (target - out).abs().mean()

# tell the optimizer what the input-output relationship is
vm = VariableManager()
vm.register(variable_name='z', shape=(128,), var_type='input')
vm.register(variable_name='target', shape(3, 256, 256), var_type='output')

# setup optimizer
opt = GradientOptimizer(model, vm, loss_fn)

# optimize
vars, out, loss = opt.optimize(num_samples=1, grad_steps=500)

detailed usage

pix2latent

pix2latent.VariableManger

class variable for managing variables. variable manager instance is initialized by

var_man = VariableManager()

pix2latent.optimizer

pix2latent.transform

pix2latent.util

pix2latent.model

pix2latent.edit