Thinking in Images

Background

Contemporary generative models primarily operate through prompt-response structures that position images as outputs of textual instruction. This approach limits the role of images to representation rather than cognition. The project questions whether latent space can instead operate as an exploratory medium where ideas evolve through iterative interaction.

Concept

The project introduces the idea of soft vector navigation within latent space. Instead of specifying precise visual targets, users provide approximate directional signals through sketching, image placement, and interaction. Meaning develops across trajectories rather than endpoints, allowing ambiguity and emergence to remain central to the process.

Process

The system experiments with iterative interaction loops between user intuition and model inference. Visual inputs act as soft constraints guiding latent transitions. Through continuous feedback, semantic neighborhoods are explored rather than selected, allowing ideas to evolve dynamically.

Reflection / Impact

Thinking in Images suggests a broader paradigm shift in how generative AI may support creative thinking. By framing latent space as a cognitive environment, the project proposes new forms of human-AI collaboration where ambiguity and exploration become productive components of reasoning.