Simple StyleGAN3 (and StyleGAN2) Generator Extension

Adds a tab to Stable Diffusion Webui that allows the user to generate images from locally downloaded StyleGAN2 or StyleGAN3 models. Created as a proof of concept and is in very early stages. This extension also provides style mixing capabilities.
NOTE: Tested only on Windows.
Based on NVlabs/stylegan3 with style mixing help from tocantrell/stylegan3-fun-pipeline

Features!

Benefits of StyleGAN

• Say goodbye to the same face; StyleGAN models are capable of generating over 4 billion diverse and highly detailed images!
• It's very fast, generating new images in tenths of a second on GPU.
• The cohesive and smooth latent space can be easily interpolated between faces, genders, animals, anything. Want to ask animal ffhq for the exact midpoint between a dog and a cat? StyleGAN can do that!
• Other cool stylegan features like DragGan, Projection of Real Images, 100% Smooth Video Transitions between two images, and more.

Features of sd-webui-gan-generator Extension

• A simple GUI tab for generating images from any StyleGAN checkpoint.
• Style mixing between two seeds for even more customization.
• Easy randomization of seed input -- let chance decide which seeds to style mix today.
• Most usefully, it can be combined with stable diffusion's inpainting/outpainting or faceswap/IP Adapter to create consistent, realistic, fictional characters!

Prerequisites (GPU Only)

Visual Studio Build Tools - WINDOWS

1. Install Visual Studio 2022: This step is required to build some of the dependencies. You can use the Community version of Visual Studio 2022, which can be downloaded from the following link: https://visualstudio.microsoft.com/downloads/
2. OR Install only the VS C++ Build Tools: If you don’t need the full Visual Studio suite, you can choose to install only the VS C++ Build Tools. During the installation process, select the option for “Desktop Development with C++” found under the “Workloads -> Desktop & Mobile” section. The VS C++ Build Tools can be downloaded from this link: https://visualstudio.microsoft.com/visual-cpp-build-tools/

Building Torch Util Libraries - WINDOWS

1. Note: When running on GPU, the following libraries are built by torch_utils: bias_act_plugin, upfirdn2d_plugin, and filtered_lrelu_plugin (for StyleGAN2 models)
2. The libraries require python3XX.lib. Because sd-webui installs into a venv, you will need to go to a locally installed Python similar to the version in sd-webui (e.g. 3.10) and manually copy the .lib yourself, as shown below:

• Open command prompt and type python
• >>> import os, sys
• >>> os.path.dirname(sys.executable)
• Navigate to indicated directory and look for the libs folder. Copy those files and create a libs folder in the sd-webui environment folder: stable-diffusion-webui\venv\scripts\libs
• Alternatively you can add Python\Python310\libs to your system path variable

Notes:

• If you get failed build, it mostly will be either due to missing library (e.g. python310.lib) or your cuda toolkit is not compatible with your GPU or pytorch installation.

Installation

1. Install AUTOMATIC1111's Stable Diffusion Webui
2. Navigate to the .\extensions and click Install from URL
3. Place https://github.com/zerxiesderxies/sd-webui-gan-generator/ under URL for extension's git repository and Install
   (Note you may need to manually add the directory name sd-webui-gan-generator)

Usage

NOTE: StyleGAN2/3 pretrained checkpoints (pickles) contain additional classes (e.g. torch_utils) that are not compatible with stable-diffusion-webui's pickle scanner. You will need to set --disable-safe-unpickle in order to load them.
TODO: Need Workaround
WARNING: Setting --disable-safe-unpickle turns off the safe pickle check and exposes sd-webui to malicious code hidden in pickle files. Use at your own risk. Please verify the integrity of your .pkl files before using.

Downloading models

1. Download any StyleGAN2 or StyleGAN3 model you prefer.

• Recommend either ffhq or celeba pre-trained networks from NVlabs
  stylegan3 checkpoints here
  stylegan2 checkpoints here

2. Place the checkpoint .pkl files in your extensions\sd-webui-gan-generator\models folder

User Interface

The Simple Image Gen tab handles basic seed to image generation.

1. Under model selection, select your model from the drop down menu. Click refresh to update the model list.
2. Under generation device, select whether you want to use your CPU cpu, Nvidia GPU cuda:0, or Mac GPU mps.
3. How to Generate an Image:

• Select a seed (integer from 0 to 2^32-1)
• Select your truncation psi (0.7 is a good value to start with).
• Click Generate Simple Image to generate the image. Note: this could take some time on a slower CPU. Check the command window for status.
• If generating first image on cuda/GPU, wait for bias_act_plugin, and filtered_lrelu_plugin to build.
• Hit the dice 🎲 button for random seed (-1 value), and the recycle ♻ button to replay the last seed.
• If you are happy with the image, you can send the seed to the Seed Mixer tab for further processing.
• Images are saved to default_output_dir. Click the folder icon 📂 to open the folder in file browser.

Seed Mixing

The Seed Mixing tab includes simple seed mixing features. Seed mixing is the process of transferring elements from one image into another. See explanation for further information.

1. Seeds imported from simple gen page, or input your Seed 1 (left) and Seed 2 (right) directly

• Can also click clover icon 🍀 to randomly populate both seed1 and seed2
• Can also click dice icon 🎲 to randomly pick seed1 and seed2 at runtime, and recycle icon to replay the last random seeds.

2. Select your truncation psi and seed mix (cross fade)
3. Select your method of style transfer in the drop-down menu.
4. Click Generate Style Mixing. You should see three images being generated: the first two seeds and the mixed image in the middle.

Explanation of the Parameters

• Seed: Integer input to create the latent vector. Each seed represents an image. Range of 32-bit unsigned integer (0 to 2^32-1).
• Truncation psi: A float that represents how much to deviate from the global average (1 = no truncation, 0 = global average). Higher values will be more diverse but with worse quality. Default is 0.7.
• Cross-fade Interpolation: A float slider for mixing the latent space between the two images.
• Transfer Method:
  • Coarse: This modifies layers 0-6, which govern high-level features. Use this if you want to modify pose, face shape, or hair style, but keep the compositional details of eyebrows, nose, eye color, and skin tone.
  • Fine: This modifies layers 7-14. Use this if you want to keep the subject's pose and shape but modify fine details such as eyebrows, nose, eye color, and skin tone.
  • Total: This mode interpolates across all layers, creating a truly linear blend between the images.

Credits

Please give credit to ZerxiesDerxies if you end up using this. Credits to NVLabs and tocantrell. This is my first gradio project and a beginner in ML/AI, so take this with a grain of salt. Mac OS support, interpolation fix, and GUI enhancements by dfl