On Nonlinear Latent Transformations for GAN-based Image Editing - PyTorch implementation

On Nonlinear Latent Transformations for GAN-based Image Editing Valentin Khrulkov, Leyla Mirvakhabova, Ivan Oseledets, Artem Babenko

Overview

We replace linear shifts commonly used for image editing with a flow of a trainable Neural ODE in the latent space.

w' = NN(w; \theta)

The RHS of this Neural ODE is trained end-to-end using pre-trained attribute regressors by enforcing

• change of the desired attribute;
• invariance of remaining attributes.

Installation and usage

Data

Data required to use the code is available at this dropbox link (2.5Gb).

Path	Description
data	data hosted on Dropbox
├  models	pretrained GAN models and attribute regressors
├  log	pretrained nonlinear edits (Neural ODEs of depth 1) for a variety of attributes on CUB, FFHQ, Places2
├  data_to_rectify	100,000 precomputed pairs (w, R[G[w]]); i.e., style vectors and corresponding semantic attributes
├  configs	parameters of StyleGAN 2 generators for each dataset (n_mlp, channel_width, etc)
└  inverses	precomputed inverses (elements of W-plus) for sample FFHQ images

To download and unpack the data run get_data.sh.

Training

We used torch 1.7 for training; however, the code should work for lower versions as well. An example training script to rectify all the attributes:

CUDA_VISIBLE_DEVICES=0 python train_ode.py --dataset ffhq \
--nb-iter 5000 \
--alpha 8 \
--depth 1

For selected attributes:

CUDA_VISIBLE_DEVICES=0 python train_ode.py --dataset ffhq \
--nb-iter 5000 \
--alpha 8 \
--dir 4 8 15 16 23 32 \
--depth 1

Custom dataset

For training on a custom dataset, you have to provide

• Generator and attribute regressor weights
• a dictionary {dataset}_all.pt (stored in data_to_rectify). It has the form {"ws": ws, "labels" : labels} with ws being a torch.Tensor of size N x 512 and labels is a torch.Tensor of size N x D, with D being the number of semantic factors. labels should be constructed by evaluating the corresponding attribute regressor on synthetic images generator(ws[i]). It is used to sample batches for training.

Visualization

Please see explore.ipynb for example visualizations. lib.utils.py contains a utility wrapper useful for building and loading the Neural ODE models (FlowFactory).

Restoring from checkpoint

import torch
from lib.utils import FlowFactory, LatentFlow
from torchdiffeq import odeint_adjoint as odeint
device = torch.device("cuda")
flow_factory = FlowFactory(dataset="ffhq", device=device)
odeblock = flow_factory._build_odeblock(depth=1)
# depth = -1 corresponds to a constant right hand side (w' = c)
# depth >= 1 corresponds to an MLP with depth layers
odeblock.load_state_dict(...)

# some style vector (generator.style(z))
w0 = ...

# You can directly call odeint
with torch.no_grad():
    odeint(odeblock.odefunc, w0, torch.FloatTensor([0, 1]).to(device))

# Or utilize the wrapper 
flow = LatentFlow(odefunc=odeblock.odefunc, device=device, name="Bald")
flow.flow(w=w0, t=1)

# To flow real images:
w = torch.load("inverses/actors.pt").to(device)
flow.flow(w, t=6, truncate_real=6)
# truncate_real specifies which portion of a W-plus vector to modify
# (e.g., first 6 our of 14 vectors)

A sample script to generate a movie is

CUDA_VISIBLE_DEVICES=0 python make_movie.py --attribute Bald --dataset ffhq

Examples

FFHQ

Bald	Goatee	Wavy_Hair	Arched_Eyebrows

Bangs	Young	Blond_Hair	Chubby

Places2

lush	rugged	fog

Citation

Coming soon.

Credits

• stylegan2-pytorch
• stylegan2-ada