Fusion Vision
Fusion Vision is an end-to-end application that focuses on providing creative control over the generative process of StyleGAN2. This application helps artists focus on their creations rather than getting familiar with code (Creative coders, links to colab notebooks below)
Features
- Generate images using 10+ models including
- human and anime faces
- abstract and modern art
- trypophobia and microscopic imgs
- imagenet and wildlife dataset
- cats, horses, cars and churches
- Fine-grained mixing of multiple seeds
- Control hand-crafted features (like haircolor, age, gender in faces) for each model
- Generate interpolating animations between images or animations with features controlled by audio
Background
Generative Adversarial Networs (GANs) can create images that are ORIGINAL in the true sense, but are hard to control. Using a mathematical technique called Principle Component Analysis, one can find such controls. Fusion Vision gives you those fine-grained controls over a StyleGAN2 model.
Inspiration
This artwork by Mario Klingemann inspired me to take up this project. (Play the video at 0.25x for some nightmare fuel).
Usage Instructions
For Artists
Yet to come
For Creative Coders
Stay tuned for notebooks
Using a custom trained model
If you've trained a StyleGAN2 model using the official NVIDIA code, convert your weights using this colab notebook
If you've trained your model using Kim Seonghyeon's code, you can skip the conversion.
Use the following notebook to do PCA on your model. Use the interactive widget in the notebook to fine-tune your components and save them
Deploy to Openshift Online
- Download the CLI from here
- Login using the CLI using the login command or use
oc login
- Use
scripts/build.sh
to delete the existing project and create a new one from the template - Copy the webhook url with secret from the Builds page and paste it your github settings
Credits
- Goku Mohandas for the summer 2020 incubator
- Erik Härkönen for the GANspace project
- Kim Seonghyeon for the pytorch implementation of StyleGAN2
- Justin Pinkney for the list of pretrained StyleGAN2 models
- Derrick Schultz for the StyleGAN2 workshop
License
The code of this repository is released under the Apache 2.0 license.