Chunkmogrify: Real image inversion via Segments

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.

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.