Ngx is an attempt at utilizing a neural network for VJing. It implements pix2pix (image-to-image translation with cGAN) as an ad-hoc next-frame prediction model that is trained with pairs of consecutive frames extracted from a video clip, so that it can generate an image sequence for an infinite duration just by repeatedly feeding frames back. It also has functionality for mixing (crossfading) two pix2pix models that gives unexpected variation and transition to generated video.
(The gif on the right is from a tweet by @chaosgroove.)
You can download a pre-built executable from Releases page. It contains pre-trained pix2pix models with Beeple's VJ clips.
| Model Name | Original clip | Vimeo link |
|---|---|---|
| Beeple 1 | FIBER OPTICAL | https://vimeo.com/238083470 |
| Beeple 2 | redgate.v1 | https://vimeo.com/106577855 |
| Beeple 3 | p-crawl | https://vimeo.com/116330714 |
| Beeple 4 | exhaust | https://vimeo.com/95513505 |
| Beeple 5 | quicksilver | https://vimeo.com/153493904 |
| Beeple 6 | MOONVIRUS | https://vimeo.com/155970396 |
| Beeple 7 | TENDRIL | https://vimeo.com/158477477 |
| Beeple 8 | shifting pains | https://vimeo.com/48818536 |
The original video clips are licensed under a Creative Commons Attribution license (CC-BY 3.0). These pre-trained models are also attributed to the original author.
Ngx can be controlled with the on-screen controller or a MIDI controller.
| Name | Description | MIDI CC |
|---|---|---|
| Mix | Crossfading between model 1 and 2 | 77 |
| Noise 1 | Noise injection (low density) | 78 |
| Noise 2 | Noise injection (medium density) | 79 |
| Noise 3 | Noise injection (high density) | 80 |
| Name | Description | MIDI CC |
|---|---|---|
| Feedback | Feedback rate | 81, 82 |
| FColor | False color effect | 83 |
| Hue Shift | Amount of hue shift | 55, 56 |
| Invert | Color inversion | 84 |
| Name | MIDI notes |
|---|---|
| Model Select 1 | 41, 42, 43, 44, 57, 58, 59, 60 |
| Model Select 2 | 73, 74, 75, 76, 89, 90, 91, 92 |
The default MIDI mapping is optimized for Novation Launch Control XL: You can change model with the track buttons and control parameters with the track faders. The 7th and 8th panning knobs are used to tweak the hue shift values.
This repository uses Git submodules to manage dependent packages. To check
these submodules out, execute git submodule init && git submodule update from
the command line, or specify --recursive option when initially cloning.
After cloning the repository, .pict (pix2pix weight data) files should be
manuyally downloaded and copied into the Streaming Assets directory because
these files are excluded from the repository to save bandwidth and storage
usage. Please download the pict file package and extract it into
Assets/StreamingAssets.
Prerequisites: Basic experience on a machine learning framework. It's recommended to have a look into the pix2pix tutorial from Machine Learning for Artists if you haven't tried pix2pix or any similar GAN model.
You can train your own pix2pix model using pix2pix-tensorflow. I personally recommend using Google Colaboratory for training. It's easy to set up, handy to use and cost effective (it's free!).
The followings are Colaboratory notebooks I used to train the default models. These notebooks are written to use Google Drive as a dataset storage.
- Video preprocess: Converts a video clip into AtoB image pairs.
- pix2pix-tensorflow colab
- pix2pix export: Exports a
.pictfile from a checkpoint of a trained model.
I used a Windows system with GeForce GTX 1070 for a live performance. It runs at around 21 fps; this might be the minimum practical spec for the system.