thepicassoproject: Picasso-inspired NFT fine art painted by a designer artificial intelligence

5 min readAug 17, 2022
Three artworks reminiscent of different Picasso periods painted by our designer artificial intelligence. From left to right: Early Years, Synthetic Cubism, and Neoclassicism.

What is thepicassoproject?

We are thepicassoproject, an innovative project that intersects fine art, neural networks, and non-fungible tokens (NFTs). This project comprises of a collection of artworks painted by a designer artificial intelligence trained on paintings from one of the most renowned, prolific, and influential artists of his time, Pablo Picasso. Our project celebrates the innovative spirit of Picasso and the power of Neural Networks. By carefully studying the artwork produced by the late prodigy, a neural network has been trained to paint and create completely original pieces of art that simultaneously pay homage to the artist, while re-imagining his genius. In contrast to exhibitions that showcase artist retrospectives, thepicassoproject capitalizes on machine learning to offer our interpretation of Picasso as a prospective. Artwork with unlimited possibilities. Our collection comprises of 10,000 stand-alone works of art, beautifully diverse, and unique by design.

Our designer artificial intelligence is learning to paint: (left to right, top to bottom) Works reminiscent of the Rose Period, Late Work, Analytical Cubism, and Blue Period.

Our mission

Our mission is to create beautifully diverse and original NFT art while simultaneously expanding the notion of generative art. What does the future of art look like to you? In this interpretation, we aim to spark curiosity, promote creativity, and showcase fine art. thepicassoproject serves as an effective platform to build on the emotions, styles, and innovation of Pablo Picasso. The majority of NFT projects emphasize redundancy, relying on minute variances to a common theme in an effort to generate diversity. We challenge this model and pioneer an NFT collection in which every single piece is truly unique. Each work of art in the collection has intrinsic value on its own. By owning a piece of the collection, you own generative art inspired by Picasso. This is fine art in NFT form. We want you to find joy in our art.

(left to right) Three works reminiscent of the periods Surrealism, Sketch work, and Neoclassicism.

How it works

Generative Adversarial Networks (GANs) are based on convolutional neural networks that can be trained in an unsupervised learning task on a given dataset. A GAN is comprised of two models, a generator neural network and a discriminator neural network. Both are trained simultaneously; the generator is trained to synthesize images and the discriminator is trained to classify images as either real (from the training dataset) or synthesized. The more often the generator creates an image that is a false negative (a synthesized image that was classified as real by the discriminator), the better the GAN performance.

GANs learn the essence of a dataset, i.e. for a set of dog images, the GAN will learn that dogs have four legs, one tail, and so forth. After successful training of the network, entirely new images can be generated that reflect the features and style of the training dataset. A more advanced application of this idea is to train a GAN on styles of painting, proportions, and even stylistic techniques of an artist. This allows for the generation of art in a similar manner.

We curated a dataset of artworks of Pablo Picasso and trained a new GAN on this data. Fig. 1 (a) shows the Fréchet Inception Distance (FID) of 50k generated images as a function of the training time. The lowest achieved FID score is FID50k = 22.07. This is comparable to the score achieved by Karras et al. on the MetFaces-U (unaligned MetFaces) dataset with FID50k = 18.75.[1]. Our FID score is an impressive value considering the diversity of our dataset, which comprises of many different types of images (not solely faces, as in the aformentioned project). However, Fig. 1 (b-e) indicate that the FID score alone cannot be used as a measure of success for the training on a diverse dataset. The FID score of the images in (b-e) is nominally similar, however, some generated images differ vastly in their appearances. Therefore, we carefully evaluated our training results by monitoring the evolution of synthesized works from a set of random vector inputs over the training process to ensure accurate generation. We finish the training of our network after 1800 GPU hours of training or 12 million images.

Figure 1: (a) Fréchet inception distance (FID) of 50k generated images as a function of the network training time in GPU hours. The FID score is based on the inception v3 convolutional neural network for image classification. Lower scores indicate a better performing GAN. (b), (c), (d), and (e) Generated images based on the same random vector input after 250, 500, 1000, and 1500 GPU training hours.

To further ensure that none of our synthesized images are identical to any works produced by Picasso or that any two synthesized works in our collection are identical, we introduce a shape metric and a color metric. Fig. 2 (a) highlights the distribution of the training dataset within the color and shape metric. Each individual point represents one artwork created by Picasso. No set of two images in the training dataset exhibits the same score within our metric, i.e. each of the training images is unique. Similarly, the distribution of the synthesized works in Fig. 2 (b) is nearly circular and no set of two images is identical. Furthermore, no set of any one synthesized image and any one training image is identical. We have shown that every artwork within our collection is truly unique and represents a novel piece of art.

Figure 2: (a) Distribution of the training dataset, i.e. paintings created by Picasso, within our shape and color metric. Each point represents one artwork. (b) Distribution of the synthesized dataset, i.e. our collection, within our metrics. Both datasets exhibit a similar distribution and each image is unique according to our metrics.

When and where to get one

We will mint our work on the Solana blockchain in Q4 ’22. The mint will happen on our website Details will follow.

The Solana blockchain is based on a hybrid protocol of proof-of-stake (PoS) and proof-of-history (PoH) validation that allows for very high transaction speeds. Furthermore, the hybrid protocol dramatically decreases the power consumption of the Solana network in comparison to the energy-intense mining process. The networks energy consumption to mint one NFT is as low as running a household refrigerator for 2 minutes or 0.001 % of the energy consumption of one Ethereum transaction.[2] The Solana ecosystem therefore provides an environmentally friendly and fast blockchain to deploy our artwork on and ensures low minting costs.


We want to give back. Our donation program will focus on contributing resources to non-profit organizations that promote art education for children. Every person has the potential to be innovative and creative and we hope to help facilitate the next generation of artists. We pledge 10 % of our sales profits to support organizations that work towards this goal. We will make Solana donations to specific organizations through Beneficiaries will be determined through a community voting process.

Social Media

Please support us on social media if you like our project :)





[1] Stylegan3: Paper. Accessed: 2022–07–10.

[2] Solana: Environment. Accessed: 2022–07–10




Pablo Picasso reborn. As a Neural Network. Fine art. All NFT — Building a community — Mint in Q4 ‘22