AASH-2017-Deep

This gallery consists of samples taken from my design studios (Learnt Smart) conducted at the 2017 (7–15 July) AA Shanghai Visiting School (Architectural Association School of Architecture, London). An 11-day long summer workshop held at the China 3D Print Museum. A series of experiments was carried out to investigate the deep learning techniques of neural style transfer at the scale and in the context of architecture and urbanism. 

animation of the iterative optimisation process
(Top) Diagram showing the simultaneous pairing of two different semantic maps and satellite imagery, one of Beijing’s Forbidden City and the other of Manhattan’s Stuy Town (a Ville Radieuse-like modern residential development). The target semantic map is used to guide the 2D generation of the new satellite imagery output as well as its 3D voxel heightfield output. (Bottom) The sequence in 2D showing the iterative optimisation processes of the semantic style transfer which generates an ‘improved’ urban satellite massing based on a loss minimization of both structure/’content’ and fabric/’style’. The neural style transfer (with semantic maps) was implemented using Lasagne/Theano in 2017.
animation of the iterative optimisation process
(Top) Diagram showing a pair of inputs consisting of the original floor plan of SANAA’s Moriyama House of 2005 in Japan and its corresponding semantic map. The target semantic map is used to guide the 2D generation of the new satellite imagery output as well as its 3D voxel heightfield output. (Bottom) The sequence in 2D showing the iterative optimisation processes of the semantic style transfer which generate an ‘improved’ floor plan based on a loss minimization of both structure/’content’ and fabric/’style’. The neural style transfer (with semantic maps) was implemented using Lasagne/Theano in 2017.
generative process (Rome)
Abstracted urban morphologies of 2 different cities (Shanghai and Rome) are made in the form of colour-coded urban bitmaps. The machine learning model uses these abstract bitmap inputs to sequentially generate larger 2D bitmaps which are further processed as 3D urban massing.
generative process (Shanghai)
A single 3D printed model with its bottom-side showing the machine learning generated urban massing of new ‘Rome’ and its top-side showing that of new ‘Shanghai’.

Leave a Reply

Read Next