What is Model Bending?

Model Bending is a creative, experimental technique inspired by digital circuit bending and glitch art, applied directly to generative AI systems. Instead of relying solely on text prompts, it intercepts a model's internal activations and layer parameters during generation to explore visual variations, create abstract results, or reveal how concepts are represented under the hood. By discovering and tweaking control points hidden in the model, you can unlock glitch aesthetics and unique visual styles that standard prompting simply cannot produce.

What is ComfyUI?

ComfyUI is a node-based interface for creating and executing AI media generation workflows, often supporting the most up-to-date image, audio, and video generation models. Its modular, node-based structure provides granular control over each stage of the generation process, from model loading and adaptation to encoding, sampling, and post-processing. It also features efficient memory management that intelligently loads/unloads models to accommodate different GPU VRAM setups. Backed by an extensive ecosystem of community-developed custom nodes and workflows, it provides a great environment for building custom AI generation pipelines.

LEARN MORE ABOUT MODEL BENDING

Who is this workshop for?

Our workshop welcomes artists, creative technologists, and AI enthusiasts who want to break away from conventional text-to-image prompting. Whether you are experienced with ComfyUI or an adventurous artist looking to push neural networks beyond their intended visual boundaries, this session is tailored for anyone interested in practice-based model exploration and experimental generative aesthetics.

What will be covered?

Participants will be guided through a hands-on walkthrough of model-bending techniques inside ComfyUI:

• Introduction to Model Bending: What are activations and how do diffusion models work? What does bending mean in this context?

• ComfyUI Setup: A brief guide to ComfyUI and installing the required custom nodes.

• Practical Applications: Applying bending techniques across Text-to-Image, Image-to-Image, and LoRA setups. We will use the CommonCanvas model, which is trained only on Creative Commons data.

• Layer Exploration: Learning how to use the plugin’s interactive interface to visualize model architecture and target specific layers.