Chunkmogrify: Real image inversion via Segments

Overview

Chunkmogrify: Real image inversion via Segments

Logo

Teaser video with live editing sessions can be found here

This code demonstrates the ideas discussed in arXiv submission Real Image Inversion via Segments.
http://arxiv.org/abs/2110.06269
(David Futschik, Michal Lukáč, Eli Shechtman, Daniel Sýkora)

Abstract:
We present a simple, yet effective approach to editing real images via generative adversarial networks (GAN). Unlike previous techniques, that treat all editing tasks as an operation that affects pixel values in the entire image in our approach we cut up the image into a set of smaller segments. For those segments corresponding latent codes of a generative network can be estimated with greater accuracy due to the lower number of constraints. When codes are altered by the user the content in the image is manipulated locally while the rest of it remains unaffected. Thanks to this property the final edited image better retains the original structures and thus helps to preserve natural look.

before after

before after

What do I need?

You will need a local machine with a relatively recent GPU - I wouldn't recommend trying Chunkmogrify with anything older than RTX 2080. It is technically possible to run even on CPU, but the operations become so slow that the user experience is not enjoyable.

Quick startup guide

Requirements:
Python 3.7 or newer

Note: If you are using Anaconda, I recommend creating a new environment to run this project. Packages installed with conda and pip often don't play together very nicely.

Steps to be able to successfully run the project:

  1. Clone or download the repository and open a terminal / Powershell instance in the directory.
  2. Install the required python packages by running pip install -r requirements.txt. This might take a while, since it will download a few packages which will be several hundred MBs of data. Some packages might need to compile their extensions (as well as this project itself), so a C++ compiler needs to be present. On Linux, this is typically not an issue, but running on Windows might require Visual Studio and CUDA installations to successfully setup the project.
  3. Run python app.py. When running for the first time, it will automatically download required resources, which are also several hundred megabytes. Progression of the download can be monitored in the command line window.

To see if everything installed and configured properly, load up a photo and try running a projection step. If there are no errors, you are good to go.

Possible problems:

Torch not compiled with CUDA enabled.
Run

pip uninstall torch
pip cache purge
pip install torch -f https://download.pytorch.org/whl/torch_stable.html

Explanation of usage

Tutorial video: click below

Open an image using File -> Image from File. There is a sample image provided to check functionality.

Mask painting:
Left click paints, right click unpaints. Mouse wheel controls the size of the brush.

Projection:
Input a number of steps (100 or 200 is ok, 500 is max before LR goes to 0 currently) and press Projection Steps. Wait until projection finishes, you can observe the global image view by choosing output mode Projection Only during this process. To fine-tune, you can perform a small number of Pivotal Tuning steps.

Editing:
To add an edit, click the double arrow down icon in the Attribute Editor on the left side. Choose the type of edit (W, S, Styleclip), the direction of the edit, and drag the sliders to change the currently masked region. Usually it's necessary to increase the multiplier before noticeable changes are reflected via the direction slider.

Multiple different edits can be composed on top of each other at the same time. Their order is largely irrelevant. Currently in the default mode, only one region is being edited, and so all selected edits apply to the same region. If you would like to change the region, you can Freeze the current image, and perform a new projection, but you will lose the ability to change existing edits.

To save the current image, click the Save Current Image button. If the Unalign checkbox is active, the program will attempt to compose the aligned face back into the original image. Saved images can be found in the SavedImages directory by default. This can be changed in _config.yaml.

Keyboard shortcuts

Current keyboard shortcuts include:

Show/Hide mask :: Alt+M
Toggle mask painting :: Alt+N

W-space editing

Source for some of the basic directions:
(https://twitter.com/robertluxemburg/status/1207087801344372736)

To add your own directions, save them in a numpy pickle format as a (num_ws, 512) or (1, 512) format and specify their path in w_directions.py.

Style-space editing (S space edits)

Source:
StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation
(https://arxiv.org/abs/2011.12799)
(https://github.com/betterze/StyleSpace)

The presets can be found in s_presets.py, some were taken directly from the paper, others I found by manual exploration. You can perform similar exploration by choosing the Custom preset once you have a projection.

StyleCLIP editing

Source:
StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery
(https://arxiv.org/abs/2103.17249)
(https://github.com/orpatashnik/StyleCLIP)

Pretrained models taken from (https://github.com/orpatashnik/StyleCLIP/blob/main/utils.py) and manually removed the decoder from the state dict, since it's not used and takes up majority of file size.

PTI Optimization

Source:
Pivotal Tuning for Latent-based Editing of Real Images
(https://arxiv.org/abs/2106.05744)

This method allows you to match the target photo very closely, while retaining editing capacities.

It's often good to run 30-50 iterations of PTI to get very close matching of the source image, which won't cause a very noticeable drop in the editing capabilities.

Attribution

This repository makes use of code provided by the various repositories linked above, plus additionally code from:

styleganv2-ada-pytorch (https://github.com/NVlabs/stylegan2-ada-pytorch)
poisson-image-editing (https://github.com/PPPW/poisson-image-editing) for optional support of idempotent blend (slow implementation of blending that only changes the masked part which can be accessed by uncommenting the option in synthesis.py)

Citation

If you find this code useful for your research, please cite the arXiv submission linked above.

Owner
David Futschik
PhD student @ CTU Prague, Czech Republic.
David Futschik
DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting

DenseCLIP: Language-Guided Dense Prediction with Context-Aware Prompting Created by Yongming Rao*, Wenliang Zhao*, Guangyi Chen, Yansong Tang, Zheng Z

Yongming Rao 322 Dec 31, 2022
Repository of continual learning papers

Continual learning paper repository This repository contains an incomplete (but dynamically updated) list of papers exploring continual learning in ma

29 Jan 05, 2023
Maximum Spatial Perturbation for Image-to-Image Translation (Official Implementation)

MSPC for I2I This repository is by Yanwu Xu and contains the PyTorch source code to reproduce the experiments in our CVPR2022 paper Maximum Spatial Pe

51 Dec 14, 2022
[AAAI-2022] Official implementations of MCL: Mutual Contrastive Learning for Visual Representation Learning

Mutual Contrastive Learning for Visual Representation Learning This project provides source code for our Mutual Contrastive Learning for Visual Repres

winycg 48 Jan 02, 2023
PromptDet: Expand Your Detector Vocabulary with Uncurated Images

PromptDet: Expand Your Detector Vocabulary with Uncurated Images Paper Website Introduction The goal of this work is to establish a scalable pipeline

103 Dec 20, 2022
Faster RCNN with PyTorch

Faster RCNN with PyTorch Note: I re-implemented faster rcnn in this project when I started learning PyTorch. Then I use PyTorch in all of my projects.

Long Chen 1.6k Dec 23, 2022
Sequence to Sequence (seq2seq) Recurrent Neural Network (RNN) for Time Series Forecasting

Sequence to Sequence (seq2seq) Recurrent Neural Network (RNN) for Time Series Forecasting Note: You can find here the accompanying seq2seq RNN forecas

Guillaume Chevalier 1k Dec 25, 2022
TensorFlow Implementation of Unsupervised Cross-Domain Image Generation

Domain Transfer Network (DTN) TensorFlow implementation of Unsupervised Cross-Domain Image Generation. Requirements Python 2.7 TensorFlow 0.12 Pickle

Yunjey Choi 864 Dec 30, 2022
PFENet: Prior Guided Feature Enrichment Network for Few-shot Segmentation (TPAMI).

PFENet This is the implementation of our paper PFENet: Prior Guided Feature Enrichment Network for Few-shot Segmentation that has been accepted to IEE

DV Lab 230 Dec 31, 2022
Source code of our work: "Benchmarking Deep Models for Salient Object Detection"

SALOD Source code of our work: "Benchmarking Deep Models for Salient Object Detection". In this works, we propose a new benchmark for SALient Object D

22 Dec 30, 2022
DeepOBS: A Deep Learning Optimizer Benchmark Suite

DeepOBS - A Deep Learning Optimizer Benchmark Suite DeepOBS is a benchmarking suite that drastically simplifies, automates and improves the evaluation

Aaron Bahde 7 May 12, 2020
Self-labelling via simultaneous clustering and representation learning. (ICLR 2020)

Self-labelling via simultaneous clustering and representation learning 🆗 🆗 🎉 NEW models (20th August 2020): Added standard SeLa pretrained torchvis

Yuki M. Asano 469 Jan 02, 2023
Collapse by Conditioning: Training Class-conditional GANs with Limited Data

Collapse by Conditioning: Training Class-conditional GANs with Limited Data Moha

Mohamad Shahbazi 33 Dec 06, 2022
NeuralDiff: Segmenting 3D objects that move in egocentric videos

NeuralDiff: Segmenting 3D objects that move in egocentric videos Project Page | Paper + Supplementary | Video About This repository contains the offic

Vadim Tschernezki 14 Dec 05, 2022
The official implementation of the Interspeech 2021 paper WSRGlow: A Glow-based Waveform Generative Model for Audio Super-Resolution.

WSRGlow The official implementation of the Interspeech 2021 paper WSRGlow: A Glow-based Waveform Generative Model for Audio Super-Resolution. Audio sa

Kexun Zhang 96 Jan 03, 2023
190 Jan 03, 2023
CARL provides highly configurable contextual extensions to several well-known RL environments.

CARL (context adaptive RL) provides highly configurable contextual extensions to several well-known RL environments.

AutoML-Freiburg-Hannover 51 Dec 28, 2022
CVPR 2022 "Online Convolutional Re-parameterization"

OREPA: Online Convolutional Re-parameterization This repo is the PyTorch implementation of our paper to appear in CVPR2022 on "Online Convolutional Re

Mu Hu 121 Dec 21, 2022
Unofficial TensorFlow implementation of Protein Interface Prediction using Graph Convolutional Networks.

[TensorFlow] Protein Interface Prediction using Graph Convolutional Networks Unofficial TensorFlow implementation of Protein Interface Prediction usin

YeongHyeon Park 9 Oct 25, 2022
DIR-GNN - Discovering Invariant Rationales for Graph Neural Networks

DIR-GNN "Discovering Invariant Rationales for Graph Neural Networks" (ICLR 2022)

Ying-Xin (Shirley) Wu 70 Nov 13, 2022