AI Hacking / Post-Generative Strategies for Diffusion Models - Einzelansicht

Generative diffusion models have become central in contemporary AI-generated media for producing highly refined images and videos from noise through iterative denoising processes. These systems are optimized for stability, and the structured nature of the latent space makes them aesthetically homogenous.

This hands-on course explores how generative AI systems, specifically diffusion models, can be disrupted and creatively misused. Participants will engage directly with the inner mechanics of these models to understand how they function and how their processes can be disrupted.  The focus moves from conventional uses of generative models such as prompt optimization, fine-tuning, and output quality to the processes, limits, and internal logic that define these systems.

The course follows a practice-based methodology in which participants carry out experiments such as injecting and manipulating different types of noise, using non-standard inputs like cross-modal signals, and exploring latent space manipulations. Together, we will investigate ways to destabilize optimization processes and rethink the role of randomness, entropy, and error in generative systems.

The implementation takes place within an open-source user interface for diffusion models. Participants engage with readings and discussions of relevant research papers and take part in practical work. Lectures are held weekly, with an intensive hacking weekend with the contribution of a software developer.

Open to students from all faculties, the course is designed to bring together participants from artistic fields such as art and design or architecture, and from technical backgrounds including computer science and HCI. Collaboration between creative and technical fields is expected, with a shared interest in experimental use of AI.

Interdisciplinarity | he course brings together approaches from media art, computer science, experimental informatics, philosophy of technology, and design-based experimentation and offers a space where technical experimentation and creative exploration inform one another.

Learning Objectives

• Understand the basic architecture and functioning of diffusion-based generative models, with a focus on visual media synthesis.
• Develop practical skills in package management for beginner-level participants and use of version control tools.
• Technical and creative fluency in developing experimental generative systems. Learn to challenge the default logic of machine learning systems.
• Work with the ComfyUI to build and customize diffusion pipelines for image/video generation. The ability to create and modify custom nodes in ComfyUI (e.g. samplers, noise modules) using Python and PyTorch for advance participants. Develop skills to noise injection, and latent space manipulation, test non-standard inputs to trigger unexpected model behaviors.
• Reflect on the aesthetic and cultural significance of generative AI through hands-on projects and creative outputs.
• Develop skills in research, teamwork, and critical analysis

Didactic Concept | The course follows a practice-based learning methodology that combines technical instruction with experimentation. Students engage with generative diffusion models through a series of structured exercises, guided experiments, and open-ended projects.

The course is structured around a combination of weekly lectures, lab sessions, student-led presentations and an intensive hacking weekend with the contribution of a software developer. Lectures introduce core concepts, lab courses offer technical instruction in tools, and student paper presentations provide a platform for individual research. The hacking weekend provides space for intense experimentation and collaborative prototyping. This session also provides real-time support for developing experimental pipelines.

The course is conducted as a „Students' Bauhaus.Module" by Funda Zeynep Ayguler (stud. MA KG). The mentorship lies withProf. Jörn Hintzer (KG). 

No formal prerequisites. The course is open to advanced bachelor and master’s students from all faculties. Students from technical disciplines are expected to be familiar with Python programming, while students from artistic fields should have an interest in experimental approaches to AI.

At the end of the course, every student, will complete an individual or small group project. Students are expected to actively participate in discussions and weekend workshop, present their ongoing experiments, contribute to the collective exhibition, and develop a final project that reflects both technical engagement and conceptual depth. The grading criteria are as follows: Attendance (10%), Presentations/Exercises (20%), Contribution to the exhibition (10%), and the Final work (60%).

The course is conducted as a „Students’ Bauhaus.Module” and open to all Master students of the faculties of Architecture and Urbanism, Civil and Environmental Engineering, Art and Design, and Media. Before registering, please consult your academic advisor and clarify whether this course can be credited to your curriculum. If required, you can conclude a learning agreement (DE/EN) before the start of the course.