Masterclass at UPDATE Festival, Liepāja
Festival: UPDATE Festival, Liepāja When: 21 and 24 September 2026, 12:00 Riga time (masterclass in two parts)
Latent Vandalism: The Joy of Productive Damage to Text-to-Image Synthesis Pipelines
Text-to-image models have evolved into sophisticated engines of template culture, trained to reproduce standardized aesthetics. Fatigued by the constant flood of polished results and the arms race for images benchmarked on visual coherence, this online masterclass explored once again the charm of AI weirdness — the failure in generative AI and the epistemic value of productive damage.
Drawing inspiration from glitch studies, participants worked directly with Diffusion Transformers (DiT), focusing on the role of embeddings in image-text correlation from embedding space to latent space back into pixel space. Through hands-on meddling with the pipeline, we systematically damaged and reconfigured the semantic substrate that guides image generation, deliberately perturbing inputs to understand the system’s sensitivity and dynamics.
This counterfactual, gently adversarial approach positions productive damage as a research method, tracing the values, design choices, and visual norms inscribed in text-to-image systems.