Workflow from input to generated output
The end-to-end pipeline, from facade input through the trained network to the generated facade design.
Pre-training database
160 sets of collected original images of Hong Kong building facades, organized for pre-calibration.
Color categories in the building label image
Elemental information of complex building facades is classified by color for later network training.
Data augmentation
Augmentation expands the training set to improve the robustness of the generated results.
The test results
Generated facades produced by the trained network across whole and local design tasks.
The training process of generator and discriminator
The adversarial training between generator and discriminator over successive iterations.
The generator and discriminator loss values
Loss curves tracking the convergence of the generator and discriminator during training.