- 15 modules: pixels, geometry, classical ML, GANs, VAEs, Diffusion Models. Pre/post test: 15.1% to 50.5% (Cohen's d = 1.615, p = .008). Dual-format pedagogy: hands-on discovery and conceptual deep-dives. TouchDesigner integration for creative coding workflows.
Open-source AI curriculum (15 modules) built as a Master's thesis, guiding learners from pixel manipulation to GANs, VAEs, and Diffusion Models. A user study with 9 participants showed knowledge scores rise from 15.1% to 50.5% (Cohen's d = 1.615, p = .008).
- 〔0〕 Python
- 〔1〕 NumPy
- 〔2〕 Jupyter
- 〔3〕 PyTorch
- 〔4〕 TensorFlow
- 〔5〕 scikit-learn
- 〔6〕 OpenCV
- 〔7〕 TouchDesigner
- Evolvable boid flocking (separation, alignment, cohesion). Predator black holes with gravitational lensing. Tournament selection and uniform crossover across generations. Quorum sensing: density-triggered collective defense.
Real-time generative art installation combining boid flocking with an evolutionary algorithm — stars evade predator black holes and pass survival traits across generations. Exhibited at the Pixels2GenAI exhibition, Berlin (2026).
- 〔0〕 Python
- 〔1〕 Pygame
- 〔2〕 NumPy
- Physarum agents sense, rotate, and deposit pheromone trails. Hebbian synapses form on agent co-activation. GPU vitals drive color, bloom, and step size. GLSL compute shaders, 60 FPS at 1920×1080.
Real-time generative artwork where live GPU vitals — temperature, clock, utilization, power — sculpt a mycelial Physarum simulation coupled to a Hebbian neural network. Up to 200,000 agents render at ~60 FPS via GLSL compute shaders.
- 〔0〕 Python
- 〔1〕 OpenGL
- 〔2〕 GLSL
- 〔3〕 CUDA
- 〔4〕 Pygame
- 〔5〕 SciPy
- ControlNet conditioned on Canny edges and MediaPipe landmarks. DDPM with linear, cosine, and quadratic noise schedules. Async generation pipeline keeping display at 30 FPS. Optional sound reactivity and ambient audio synthesis.
Interactive art installation where visitors watch themselves dissolve into noise and crystallize as anime artwork via Stable Diffusion and ControlNet. Exhibited at IT Studio Academis, Berlin (2026), with async generation at ~3s/frame and live display at 30 FPS.
- 〔0〕 Python
- 〔1〕 Stable Diffusion
- 〔2〕 ControlNet
- 〔3〕 MediaPipe
- 〔4〕 OpenCV