MODULE MAP

Join the Nova Collective and discover the fundamental difference between classifying data and creating it. You'll learn probability distributions, sampling, and noise — the raw materials of generative AI — using concepts you already know from frontend creative coding.

Explore the hidden dimensions where generative models dream. You'll build an interactive latent space explorer, learning how data is compressed into compact representations and how walking through latent space creates smooth transformations — concepts that mirror URL encoding, CSS transitions, and component props.

Pit two neural networks against each other in a creative duel — a generator that forges art and a discriminator that critiques it. You'll build both networks, train them in an adversarial dance, and tackle the infamous mode collapse problem — using concepts from testing, validation, and load balancing.

Move beyond GANs to Variational Autoencoders — models that learn smooth, structured latent spaces for controllable generation. You'll master the reparameterization trick, KL divergence, and creative VAEs — concepts that map to compression pipelines, seeded randomness, and design system tokens.

First high-quality images from MUSE — but at 43 seconds per render, the gallery kiosks grind to a halt. Erdem panics. You'll master the diffusion process from noise schedules to classifier-free guidance, then use progressive distillation to bring generation down to 3.4 seconds. Diffusion is just progressive image loading — the blur-up technique you already know.

Yara asks the question that changes everything: "Can the AI paint a photograph in the style of Monet? Of Basquiat? Of Yara?" You'll separate content from style, build neural style transfer pipelines, and run them in real time in the browser. Style transfer is CSS theme switching — swapping visual identity while preserving underlying structure.

The midpoint breakthrough. MUSE generates its first image from the text prompt "a sunset over a digital city." Rough, but recognizable. You'll build models that understand multiple data types — vision-language models, cross-modal embeddings, image captioning, and text-to-image generation. Multimodal is full-stack: frontend, backend, and database working as one system.

CLIP integration gives MUSE nuance — "a melancholy cityscape at twilight" versus "a city at night" produce meaningfully different images. Rohan says mood is a direction in a 512-dimensional vector space. But technically perfect output can be artistically dead. You'll master contrastive learning, CLIP architecture, zero-shot classification, and prompt engineering — the Algolia multi-index search of AI.

The gallery walls pulse with generative visuals, but the room is silent. Yara declares sound is not decoration — it is a dimension. Rohan reveals the secret: a spectrogram is an image. You already know how to generate images from latent spaces. Audio is the same math, different domain. You'll transform waveforms into tensors, synthesize sounds with neural networks, and build interactive soundscapes using the Web Audio API you already know.

The Biennale is three weeks away. Erdem runs the numbers: 200 people generating simultaneously, zero lag tolerance, gallery Wi-Fi that drops packets like confetti. You'll profile generation bottlenecks, optimize models with quantization and distillation, implement caching strategies with Service Workers, handle concurrent WebSocket connections, and architect a physical-digital installation as a PWA.

Opening night. 247 concurrent visitors at peak. A child whispers "a dragon made of stars" and Nova Canvas generates it in real time. Everything the collective has built — image generation, style transfer, text-to-image, audio synthesis, multimodal alignment — comes together in a single installation. You'll integrate all modalities, design visitor experiences, load-test for opening night, and keep it running.