“What would happen if you resurrected five artists who lived with 500 years apart?”, he wonders Robert Luciani. I’d say probably a good dinner, and if you can get them to like each other with the help of a few good bottles of wine, they might eventually collaborate on making ‘The impossible statue’, an impressive stainless steel figure of 1.5 meters and a half ton that fuses their styles and is now exhibited in the Tekniska Museumthe Swedish National Museum of Science and Technology in Stockholm.
Omar Kardoudi
Luciani is a software engineer at The AI Framework, the AI and data consultancy that created the AI models and workflow that led to the creation of this work of art. The process was an enormous challenge, from the creation of models to manufacturing and final assemblysays Sandvikthe Swedish engineering company that built the statue to showcase their CNC manufacturing capabilities.
An almost impossible challenge
The first challenge, according to Sofia Sirvell, Sandvik Group’s digital director, was to create a 3D model that represented these artists. In an email interview, Sirvell tells me that the statue was designed by training multiple AI models on the work of five of the world’s greatest and most renowned sculptors, balancing some of their best-known attributes: “We incorporated the dynamic, unbalanced poses of the Italian artist Miguel Angel and the musculature and reflection of the French sculptor auguste rodin. We are also inspired by the expressionist sentiment of the German artist Käthe Kollwitzand in the approach of the impulse and the mass of the Japanese poet and sculptor Takamura Kotaro. This collaboration was completed with influences from the African-American sculptor Augusta Savageinspiring us in the rebellion that their figures reflect”.
Although today exists certain 3D generative AI technology, is still quite crude, so they had to use 2D tools. Initial work began in August 2022, he explains. At that time, stable diffusion (SD) had a very low image continuity, that is, the ability to maintain the coherence of the generated images with each other. To overcome this limitation, they trained an SD model on a relatively small dataset of statues synthesized from 3D images. However, this model did not have an understanding of the artists they wanted the AI to become. For this, they used midjourney and DALL-E to synthesize poses and themes from each artist, which were fed into the Stable Diffusion model to produce the final 2D image.
“We set up an automated system that combined the capabilities of the three AI tools we had chosen to create our initial image,” Sirvell explains. “Producing consistent and definitive results in early versions of generative AI is a difficult task, especially when trying to balance conflicting artistic styles such as realism and expressionism. That’s why we use more than one tool: each one has a different strength.”
While DALL-E has a deep knowledge of the semantic and syntactic content and understands the language very well; Midjourney’s forte is interpretation of art style and more subjective attributes like ambience, emotion and style, while Stable Diffusion excels at producing high-quality imagery, he tells me. They repeated the generation process over and over again, obtaining images until they found one that they considered fulfilled the objective and could be manufactured by CNC machining.
From 2D to 3D…
Translating 2D images into a 3D model was just as difficult. Sirvell tells me that creating the point cloud and the subsequent solid model that could be fabricated required another set of AI tools. Some of these tools required more custom training to achieve maximum fidelity, such as the Midas depth estimatoran AI algorithm capable of estimating a 3D model from a monocular planar image who estimated the depth value of the statue. This initial 3D model of the body was later refined with human posture estimators, who also interpreted flat images of the human figure. Another step was the generation of the realistic fabric from which the body emerges, for which AI algorithms used in video game. Lastly, Sirvell says they applied other specialized AI software to reintroduce additional fine details lost in the previous steps.
The development of AI image generators is now so rapid that there are several generators that are already capable of performing this process autonomously, unlike when The Impossible Statue began its creative journey. “Many of these models have been vastly improved since the project began, but even today creating an equivalent 3D mesh would require multiple custom models and remains a task that is not trivial“Sirvell says.
…and to the real world
With the final 3D point cloud model in hand, they needed to bring it into the real world, a huge challenge in itself due to the size and complexity of the sculpture. “Our team used Mastercam computer aided manufacturing softwareresulting in a robust model with more than nine million polygons and very intricate details.
However, the statue could not be made from a single piece of steel. Fabrication had to first be simulated on a digital twin before beginning any CNC (computer numerical control) sculpting process, which uses drills to mold the steel blocks to the 3D mesh. To do this, a program called Vericuta software developed by Sandvik that simulates and optimizes the manufacturing process so that you use the least amount of steel possible and the process is safe and efficient. According to the company, “not a single piece of the statue had to be scrapped or redone throughout the process, as each had been digitally perfected before machining began.”
After three simulations, your software recommended the creation of 17 separate pieces that had to go through various processes of turning, milling and drilling, to then be perfectly assembled into the final shape. The process gave rise to 40 million lines of G code, the program that guides the machines to sculpt the pieces. The company claims that this entire process halved the amount of steel needed. Sandvik also ensures that the accuracy of the finished work—as measured by its laser tools and software—is only 30 microns from the original 3D modelwhich “makes the intersections between the different pieces practically invisible to the naked eye“.
The company states that “this level of precision is a remarkable feat considering its size and the complexity of the 3D model.” It certainly looks like an impressive manufacturing feat. But what really matters is how everything was done from conception to completion. How an AI-based creation process can be guided by humans and work with machines at various levels to create something tangible, not just an image, is a shiny sign of what these technologies can bring us, like the AI-generated alien NASA spacecraft parts we wrote about a few months ago. awaits us a Brilliant Futureif we manage to keep the right course.
“What would happen if you resurrected five artists who lived with 500 years apart?”, he wonders Robert Luciani. I’d say probably a good dinner, and if you can get them to like each other with the help of a few good bottles of wine, they might eventually collaborate on making ‘The impossible statue’, an impressive stainless steel figure of 1.5 meters and a half ton that fuses their styles and is now exhibited in the Tekniska Museumthe Swedish National Museum of Science and Technology in Stockholm.
