nonstandardstudio

The integration of AI into architecture is a game-changer, and I am confident in its potential to usher in groundbreaking changes. From Creative AI, creating intelligent design to optimizing building performance, AI promises to redefine the entire architectural process.

For those familiar with my research, I've been advocating for a long time for the development of industry-specific neural networks. Since 2017, I have been working on the DeepHimmelblau project in collaboration with Coop Himmelblau. Our objective is to enhance design processes and augment designers' creativity.

DeepHimmelblau is a node-based system consisting of multiple neural networks tailored to various design tasks (Organizational – Technical – Gestalt). The node-based structure enables task-dependent strategies of interconnected nodes to be established in response to discrete design tasks, specific design problems or the nature of any given investigation. The network not only allows the connection of multiple nodes, but also permits the semantic levels of the various node network layers to be combined, blended, and swapped between network nodes. One notable addition is the diffusion node, which was trained on CoopHimmelblau's proprietary image and text-based dataset. This node surpasses the limitations and aesthetic preferences of generic models, such as MJ, Stable Diffusion, and Dalle2.

Below is a collection of generated comparison outputs.

prompt:
"an interior circulation of a building made out of concrete, sinuous lines, aluminum triangular panelization, by CoopHimmelblau"

Here you can find more information about the project, and a roundtable discussion where I introduced some of the latest features of the DeepHimmelblau, such as establishment of coherent compositional relationships across interior, exterior, and elevations, ensuring compositional consistency and accurate approximation of different views of the same design.

https://onlinelibrary.wiley.com/doi/epdf/10.1002/ad.2808

https://youtu.be/jjUb48f4ROc?t=2883

Looking back on my Gaudi+NeuralNetworks, I can’t help but feel a genuine sense of amazement at the quality of the output that the generative adversarial networks (GANs) I developed achieved. While GANs are known to often produce mushy or flawed images, the networks developed for this project output results close to the quality of current diffusion models like , in terms of resolution and composition coherence.

I made some great improvements to the network over the past week. Looking into the development of generative networks capable of learning relevant semantic features.

3D Domain Translation using Cycle-Consistent Adversarial Networks

Some of the algorithms I started developing while at I.sd Institute of Structure and Design, Innsbruck.

The 3D Domain Translation model starts to show some promising results. Although still more work is needed to be done to allow for a better disentanglement of features.

This approach builds on the work of Jun-Yan Zhu, Taesung Park, Philip Isola, Alexei A. Efros from Berkeley Ai Research (Bair) Laboratory, UC Berkeley

More detailed description soon...

@ Miami, Florida