Composicion Constructiva by J. Torres Garcia

The Uruguayan painter Joaquin Torres Garcia is particularly renowned for introducing Constructivism to South America. Torres-García’s style incorporates elements of Cubism, which flattened perspective and emphasized the two-dimensional nature of the canvas. His style also includes Constructivism and neoplasticism, where he aimed to create a universal language in art that could transcend cultural and regional differences. His mature style, known as Universal Constructivism, is marked by a balance between abstraction and figuration, using a gridded composition filled with symbolic elements ranging from fish to stars, combining cosmic, historical, and personal iconography. His work is currently being displayed in the Museo Reina Sofia in Spain, the Guggenheim, and the MoMa, among others.

A Collection Destroyed by Fire and AI Outpainting.

In 1978 a fire destroyed several pieces of the painter’s collection in the Museum of Modern Art in Brazil. Some pieces were rescued albeit severely damaged, like the one depicted below. The role of Artificial Intelligence is to imagine possible reconstrucions for the damaged work that can be truthful to the paint. The algorithm learns the painter’s patterns and style through only scanning through the surviving painting without having prior knowledge of the original undamaged work.

Generating Inpaited Candidates

The challenge for the AI model is to learn to reproduce the style of a somehow niche paintor without having prior information. Additionally, the algorithm needs to understand the separation between the healthy and damaged parts of the painting to overpaint only on the burnt surface leaving the remaining healthy parts intact. Overall, the AI produced three high-quality outpainting candidates that match Torres Garcia’s style very well, as can be seen below.

Selecting the Best Inpainting Candidate

As shown above, the model can generate several high-quality candidates suited for the outpainting of the piece, withouth having seen the original work. The last step consisted in selecting ona final result that best follows Torres Garcia’s style. Our model’s selection process is based on ultimately choosing the restoration outcome that most authentically resonated with the author, ensuring the final result shows as much resemblance as possible with a true painting.

Related Publications

The Artificial Intelligence model is described in our CVPR 2022 publication. The candidate selection model is described in our ICML 2022 publication.