GAN-generated architectural sculptural relief

A dataset of fragmented and decontextualized Greco-Roman sculptural relief underlies the generation of uncanny forms that straddle the unrecognizable and the familiar. Samples include those drawn from the Pergamon Altar: a Greek construction originating in modern-day Turkey, disassembled in the late 19th century, and re-assembled in the early 20th century in a Berlin museum. The project operates similarly. It begins with a disassembly of selected sculptural forms into fragments that can be described as deformations of a flat sheet.

Where ML processes often struggle to describe three-dimensional form, these "vector displacement maps" are comprehensible to the machine, and serve to train a neural network - a gently modified implementation of StyleGAN - to understand the form-language of the selected source material. Recalling the rhythmic symmetry of frieze patterns found in traditional Western ornament, a "walk" through the latent space of Greco-Roman sculptural forms is aggregated across a surface in high relief.



Artificialis Rilievo was presented at the 2021 Venice Architecture Biennale as part of the X Venice Italian Virtual Pavillion. This work was conceived and directed by Professor Kyle Steinfeld.

A Pipeline for Representing 3d Sculptural Relief as Raster Data

1. Pergamon Altar Meshes
We used 3D scanned parts of the frieze of the Pergamon Altar, a ancient Greek temple. These sample polygon meshes acted as the initial dataset.


2. Fragments
Smaller fragments from the large mesh are selected and isolated for the dataset. The geometry is cleaned manually to be optimal for the next steps.


3. Geometry to Plane
The fragments geometry is “squashed” onto a plane, with displacements between points on the plane and locations on the 3d mesh stored as vectors separated into their x,y, and z components.


4. Vector Displacement Maps (VDM)
This vector information is stored as the RGB channels of a raster image. (32-bit TIFFs)


5. VDMs as training dataset
VDM format is both amenable to a GAN, and is able to be re-interpreted as vector displacements from a base raster plane. We used them to train a generative adversarial network (StyleGAN2-ADA) on VDM data.


6. VDMs to sculptural forms
Finally we use the synthetic data to translate VDMs “back” into pseudo-3d sculptural forms. The resulting synthetic VDMs from the GAN pipeline is used to feed the final artificial reliefs.

Training Data

Vector Displacement Map of Double Hex

Fabrication

Casting

The latent space walk was modeled in defined modular tiles. The tiles were 3D printed in a SLA printer using a special PLA-type filament, suitable for investment casting. The printed tiles were casted in bronze using a “lost-wax process”.

Bronze Post Processing

Finishing steps included the typical clipping, sanding, and de-burring, as well as the welding of brass mounting plates, and the application of a cold patina through a series of chemical baths and manual scrabbing and cleaning.

---

Team

---

Tools
Rhinoceros, Grasshopper, Adobe Creative Suite, StyleGAN, OMAX Waterjet Cutter, 3D printers, Casting methods