Home

Awesome

Hijack-GAN: Unintended-Use of Pretrained, Black-Box GANs

Pytorch 1.7.0 cvxpy 1.1.11 tensorflow 1.14

In this work, we propose a framework HijackGAN, which enables non-linear latent space traversal and gain high-level controls, e.g., attributes, head poses, and landmarks, over unconditional image generation GANs in a fully black-box setting. It opens up the possibility of reusing GANs while raising concerns about unintended usage.

[Paper (CVPR 2021)][Project Page]

Prerequisites

Install required packages

pip install -r requirements.txt

Download pretrained GANs

Download the CelebAHQ pretrained weights of ProgressiveGAN [paper][code] and StyleGAN [paper][code], and then put those weights in ./models/pretrain. For example,

pretrain/
├── Pretrained_Models_Should_Be_Placed_Here
├── karras2018iclr-celebahq-1024x1024.pkl
├── karras2019stylegan-celebahq-1024x1024.pkl
├── pggan_celebahq_z.pt
├── stylegan_celebahq_z.pt
├── stylegan_headpose_z_dp.pt
└── stylegan_landmark_z.pt

Quick Start

Specify number of images to edit, a model to generate images, some parameters for editting.

LATENT_CODE_NUM=1
python edit.py \
    -m pggan_celebahq \
    -b boundaries/ \
    -n "$LATENT_CODE_NUM" \
    -o results/stylegan_celebahq_eyeglasses \
    --step_size 0.2 \
    --steps 40 \
    --attr_index 0 \
    --task attribute \
    --method ours

Usage

Important: For different given images (initial points), different step size and steps may be considered. In the following examples, we provide the parameters used in our paper. One could adjust them for better performance.

Specify Number of Samples

LATENT_CODE_NUM=1

Unconditional Modification

python edit.py \
    -m pggan_celebahq \
    -b boundaries/ \
    -n "$LATENT_CODE_NUM" \
    -o results/stylegan_celebahq_smile_editing \
    --step_size 0.2 \
    --steps 40 \
    --attr_index 0\
    --task attribute

Conditional Modification

python edit.py \
    -m pggan_celebahq \
    -b boundaries/ \
    -n "$LATENT_CODE_NUM" \
    -o results/stylegan_celebahq_smile_editing \
    --step_size 0.2 \
    --steps 40 \
    --attr_index 0\
    --condition\
    -i codes/pggan_cond/age.npy
    --task attribute

Head pose

Pitch

python edit.py \
    -m stylegan_celebahq \
    -b boundaries/ \
    -n "$LATENT_CODE_NUM" \
    -o results/ \
    --task head_pose \
    --method ours \
    --step_size 0.01 \
    --steps 2000 \
    --attr_index 1\
    --condition\
    --direction -1 \
    --demo

Yaw

python edit.py \
    -m stylegan_celebahq \
    -b boundaries/ \
    -n "$LATENT_CODE_NUM" \
    -o results/ \
    --task head_pose \
    --method ours \
    --step_size 0.1 \
    --steps 200 \
    --attr_index 0\
    --condition\
    --direction 1\
    --demo

Landmarks

Parameters for reference: (attr_index, step_size, steps) (4: 0.005 400) (5: 0.01 100), (6: 0.1 200), (8 0.1 200)

CUDA_VISIBLE_DEVICES=0 python edit.py \
    -m stylegan_celebahq \
    -b boundaries/ \
    -n "$LATENT_CODE_NUM" \
    -o results/ \
    --task landmark \
    --method ours \
    --step_size 0.1 \
    --steps 200 \
    --attr_index 6\
    --condition\
    --direction 1 \
    --demo

Generate Balanced Data

This a templeate showing how we generated balanced data for attribute manipulation (16 attributes in our internal experiments). You can modify it to fit your task better. Please first refer to here and replace YOUR_TASK_MODEL with your own classification model, and then run:

NUM=500000
CUDA_VISIBLE_DEVICES=0 python generate_balanced_data.py -m stylegan_celebahq \
    -o ./generated_data -K ./generated_data/indices.pkl -n "$NUM" -SI 0 --no_generated_imgs

Evaluations

TO-DO

Acknowledgment

This code is built upon InterfaceGAN