Lost Memories from Latent Space

This generative art project explores the latent space and internal mechanisms of neural networks. Using a vast dataset of Renaissance Mythology paintings, CLIP-based semantic clustering is applied to divide the corpus into 100 conceptual categories. Each category informs the training of a conditional StyleGAN2 model, resulting in 100 distinct generative loops, each a traversal through latent space that visually unfolds the model’s inner structure.

At the core of each piece is a dynamic journey through the intermediate layers of the StyleGAN architecture, revealing the evolution of internal neural representations as the final image takes shape. Feature activations - local ‘blobs’ of information - are tracked across layers and channels to visualize their emergence, transformation, and interconnection. These clusters are analyzed and highlighted using color palettes sampled from the statistical distribution of the generated outputs, producing responsive heatmaps and abstract diagrams that echo the tones of the source material.

Occasional full-layer visualizations offer rare windows into the hidden logic of the network. Subtle feedback effects enhance temporal continuity, binding the technical and the aesthetic into a coherent flow. Complementing the visuals, sound is generated by translating blob features into waveforms, modulated by parameters such as cluster size and visualization type, offering a synesthetic immersion into the unfolding logic of generative synthesis.