The Code: Our implementation can be found at this Github repo: https://github.com/libreai/neural-painters-x
I am sure you know Bob Ross and his program The Joy of Painting where he taught thousands of viewers how to paint beautiful landscapes with a simple and fun way, combining colors and brushstrokes, to achieve great results very quickly. Do you remember him teaching how to paint a pixel at the time? of course not!
However, most of the current generative AI Art methods are still centered to teach machines how to ‘paint’ at the pixel-level in order to achieve or mimic some painting style, e.g., GANs-based approaches and style transfer. This might be effective, but not very intuitive, specially when explaining this process to artists, who are familiar with colors and brushstrokes.
At Libre AI, we have started a Creative AI initiative with the goal to make more accessible the advances of AI to groups of artists who do not necessarily have a tech background. We want to explore how the creative process is enriched by the interaction between creative people and creative machines.
As first step, we need to teach a machine how to paint. It should learn to paint as a human would do it: using brushstrokes and combining colors on a canvas. We researched the state-of-the-art and despite the great works, there was not really a single paper that satisfied our requirements, until we found Neural Painters: A Learned Differentiable Constraint for Generating Brushstroke Paintings by Reiichiro Nakano . This finding was quite refreshing.
Neural Painters  are a class of models that can be seen as a fully differentiable simulation of a particular non-differentiable painting program, in other words, the machine “paints” by successively generating brushstrokes (i.e., actions that define a brushstrokes) and applying them on a canvas, as an artist would do.
These actions characterize the brushstrokes and consist of 12-dimensional vectors defining the following variables:
- Start and end pressure: pressure applied to the brush at the beginning and end of the stroke
- Brush size: radius of the generated brushstroke
- Color: the RGB color of the brushstroke
- Brush coordinates: three Cartesian coordinates on a 2D canvas, defining the brushstroke’s shape. The coordinates define a starting point, end point, and an intermediate control point, constituting a quadratic Bezier curve
A tensor with actions looks like this example:
[0.7016, 0.3078, 0.9057, 0.3821, 0.0720, 0.7956, 0.8851, 0.9295, 0.3273, 0.8012, 0.1321, 0.7915],
[0.2864, 0.5651, 0.5099, 0.3430, 0.2887, 0.5044, 0.0394, 0.5709, 0.4634, 0.8273, 0.1056, 0.1702],
and these are a sample of some of the brushstrokes in the dataset:
The goal of the Neural Painter is to translate these vectors of actions into brushstrokes on a canvas. The paper explores two neural architectures to achieve such translation, one based on a variational autoencoder (VAE) and the second one based on a generative adversarial network (GAN), with the GAN-based Neural Painter (Figure 1) achieving better results in terms of quality of the generated brushstrokes. For more details please refer to the paper  .
Tinkering with Neural Painters
The code available to reproduce the experiments is offered by the author in a series of Google’s Colaboratory notebooks available in this Github repo and the dataset used is available in Kaggle. The implementation uses TensorFlow, which is great in terms of performance, but let’s face it, it is not great fun to digest TensorFlow’s code (specially without Keras ;) ).
Teaching machines is the best way to learn Machine Learning — E. D. A.
We played around with the notebooks provided, they were extremely useful to understand the paper and to generate nice sample paintings, but we decided that in order to really learn and master Neural Painters, we needed to experiment and reproduce the results of the paper with our own implementation. To this end, we decided to go with PyTorch and fast.ai as deep learning frameworks instead of TensorFlow as the paper’s reference implementation, to do some tinkering and in the process, hopefully, come with a more accessible piece of code.
Learning Neural Painters Faster
GANs are great generative models but they are known to be notoriously difficult to train, specially due to requiring a large amount of data, and therefore, needing large computational power on GPUs. They require a lot of time to train and are sensitive to small hyperparameter variations.
We indeed tried first a pure adversarial training following the paper, but although we obtained some decent results with our implementation, in terms of brushstrokes quality, it took a day or two to get there with a single GPU using a Colaboratory notebook and the full dataset.
To overcome these known GANs limitations and to speed up the Neural Painter training process, we leveraged the power of Transfer Learning
Transfer learning is a very useful technique in Machine Learning, e.g., the ImageNet models trained as classifiers, are largely used as powerful image feature extractors, in NLP, word embeddings, learned unsupervised or with minimal supervision (e.g., trying to predict words in the same context), have been very useful as representations of words in more complex language models. In Recommender Systems, representations of items (e.g., book, movie, song) or users can be learned via Collaborative Filtering and then use them not only for personalized ranking, but also for adaptive user interfaces. The fundamental idea, is to learn a model or feature representation on a task, and then transfer that knowledge to another related task, without the need to start from scratch, and only do some fine-tuning to adapt the model or representation parameters on that task.
More precisely, since GANs main components are the Generator and Critic the idea is to pre-train them independently, that is in a non-adversarial manner, and do transfer learning by hooking them together after pre-training and proceed with the adversarial training, i.e., GAN mode. This process has shown to produce remarkable results  and is the one we follow here.
The main steps are as described as follows:
(1) Pre-train the Generator with a non-adversarial loss, e.g., using a feature loss (also known as perceptual loss)(2) Freeze the pre-trained Generator weights(3) Pre-train the Critic as a Binary Classifier
(i.e., non-adversarially) using the pre-trained Generator (in evaluation mode with frozen model weights) to generate `fake` brushstrokes. That is, the Critic should learn to discriminate between real images and the generated ones. This step uses a standard binary classification loss, i.e., Binary Cross Entropy, not a GAN loss(4) Transfer learning for adversarial training (GAN mode): continue the Generator and Critic training in a GAN setting. Faster!
More in detail:
(1) Pre-train the Generator with a Non-Adversarial Loss
The training set consists of labeled examples where the input corresponds to an action vector and the corresponding brushstroke image to the target.
The input action vectors go through the Generator, which consists of a fully-connected layer (to increase the input dimensions) and of a Deep Convolutional Neural Network connected to it.
The output of the Generator is an image of a brushstroke. The loss computed between the images is the feature loss introduced in  (also known as perceptual loss ). The process is depicted in Figure 1.
(2) Freeze the pre-trained Generator
After pre-training the Generator using the non-adversarial loss, the brushstrokes look like the ones depicted in Figure 2. A set of brushstrokes images is generated that will help us pre-train the Critic in the next step.
(3) Pre-train the Critic as a Binary Classifier
We train the Critic as binary classifier (Figure 3), that is, the Critic is pre-trained on the task of recognizing true vs generated brushstrokes images (Step (2)).
We use is the Binary Cross Entropy as binary loss for this step.
(4) Transfer Learning for Adversarial Training (GAN mode)
Finally, we continue the Generator and Critic training in a GAN setting as shown in Figure 4. This final step is much faster that training the Generator and Critic from scratch as a GAN.
One can observe from Figure 2 that the pre-trained Generator is doing a decent job learning brushstrokes. However, there are still certain imperfections when compared to the true strokes in the dataset.
Figure 5 shows the output of the Generator after completing a single epoch of GAN training, i.e., after transferring the knowledge acquired in the pre-training phase. We can observe how the brushstrokes are more refined and, although slightly different to the true brushstrokes, they have interesting textures, which makes them very appealing for brushstrokes paintings.
From Brushstrokes to Paintings
Once the Generator training process is completed, we have a machine that is able to translate vectors of actions to brushstrokes, but how do we teach the machine to paint like a Bob Ross’ apprentice?
To achieve this, the Neural Painters paper  introduces a process called Intrinsic Style Transfer, similar in spirit to Neural Style Transfer  but which does not require a style image. Intuitively, the features of the content input image and the one produced by the Neural Painter should be similar.
To implement the process we freeze the Generator model weights and learn a set of action vectors that when input to the Generator will produce brushstrokes, that once combined, will create a painting given an input content image.The image features are extracted using a VGG16  network as a feature extractor, denoted as CNN in Figure 6, which depicts the whole process.
Note that the optimization process is targeted to learn the tensor of actions, while the remaining model weights are not changed, that is, the ones of the Neural Painter and CNN models. We use the same Feature Loss as before .
Finally, given an input image for inspiration, e.g., a photo of a beautiful landscape, the machine is able to create a brushstroke painting for that image :) .
- For blending the brushstrokes, we follow a linear blending strategy to combine the generated strokes in a canvas, this process is described in detail in a very nice post titled Teaching Agents to Paint Inside Their Own Dreams also by Reiichiro Nakano . We are currently exploring an alternative process that uses the alpha channel for blending.
We would like to thank Reiichiro Nakano for helping us clarifying doubts during the implementation of our Neural Painters and for his supportive and encouraging comments and feedback. Thanks a lot Reiichiro! [@reiinakano].
 Neural Painters: A Learned Differentiable Constraint for Generating Brushstroke Paintings. Reiichiro Nakano
arXiv preprint arXiv:1904.08410, 2019.
 Decrappification, DeOldification, and Super Resolution. Jason Antic (Deoldify), Jeremy Howard (fast.ai), and Uri Manor (Salk Institute) https://www.fast.ai/2019/05/03/decrappify/ , 2019.
 Fast.ai MOOC Lesson 7: Resnets from scratch; U-net; Generative (adversarial) networks. https://course.fast.ai/videos/?lesson=7 ; Notebook: https://nbviewer.jupyter.org/github/fastai/course-v3/blob/master/nbs/dl1/lesson7-superres.ipynb [Accessed on: 2019–08]
 Perceptual Losses for Real-Time Style Transfer and Super-Resolution
Justin Johnson, Alexandre Alahi, Li Fei-Fei https://arxiv.org/abs/1603.08155 , 2016
 Teaching Agents to Paint Inside Their Own Dreams. Reiichiro Nakano.
https://reiinakano.com/2019/01/27/world-painters.html , 2019
 A Neural Algorithm of Artistic Style. Leon A. Gatys, Alexander S. Ecker, Matthias Bethge. https://arxiv.org/abs/1508.06576, 2015
 Very Deep Convolutional Networks for Large-Scale Image Recognition. Karen Simonyan, Andrew Zisserman. https://arxiv.org/abs/1409.1556, 2014